The Ontology Imperative - Building Trustworthy Agentic AI: The Crisis

2d – The CDO at the Crossroads: Where AI Works, Where It Breaks, and Why the Role Most Prepared to Close the Gap Is Most at Risk

Frédéric Verhelst — Thu, 26 Mar 2026 07:06:17 GMT

Reading time: ~25 minutes

Summary

Chief Data Officers are the most prepared executives for agentic AI and the most likely to be eliminated before they can deliver. The difference is mandate. Pipelines are what agents will automate. The semantic layer is what agents cannot function without. This article explains where AI delivers and where it fails, why monitoring cannot close the governance gap, and the five questions every board must ask to decide which CDO it has.

Board Brief: The Paradox, the Frontier, and the Framework

The Chief Data Officer sits alone at the extreme upper-right of the Capgemini Research Institute's 2026 executive analysis: highest on both role transformation impact and preparedness to navigate it. The CIO and CTO share that quadrant but trail significantly on transformation impact. The CFO, COO, CEO, and the rest of the C-suite are clustered in the lower half.

The same population averages thirty months of tenure. Fifty-three point seven percent serve fewer than three years. Twenty-nine percent question whether the role should exist at all. Gartner projects that seventy-five percent of Chief Data and Analytics Officers not perceived as essential to AI success will lose their C-level position by 2027.

The role most ready for this transformation is the role most likely to be eliminated before it can deliver it.

This is not a talent problem. It is a structural verdict. And understanding that verdict is the precondition for everything that follows.

Four questions for your next board meeting:

Does your CDO hold authority over the semantic layer that agents will reason over, or only over the data pipelines agents will eventually automate?
Can your board describe, in concrete terms, the constraints governing each autonomous agent currently running under your corporate authority?
When your AI agents fail, not if, do you have the accountability structure to explain to regulators, customers, and shareholders exactly what happened and why?
Is your CDO being measured on what agents will automate, or on what agents cannot function without?

If any of these questions exposes a gap, the governance architecture discussed in Part 1 and Part 2 of this series has not yet reached the organizational level where it needs to live. The five-question board framework at the end of this article provides the diagnostic.

Executive Brief

Agentic AI works. The organizations that will capture its value are the ones that treat governance as an organizational function before it is a technical one.

I have spent twenty-five years at the intersection of data, autonomous systems, and industrial operations. I worked with BDI multi-agent systems in 2007, when formal ontologies were the only architecture that made agent behavior auditable across shifting operational contexts in a subsea drilling environment. I led the IOHN program, a semantic interoperability initiative across twenty-two energy companies on the Norwegian Continental Shelf. I was CEO of the Knowledge Graph Alliance, whose founding members included Airbus, Michelin, Bosch, and TotalEnergies. And I spent years as Head of Data Office, an experience that shaped my view of the CDO role from the inside.

In 2008 I presented at the W3C Semantic Web workshop in Houston, arguing that autonomous agents at industrial scale required formal ontological constraints before deployment. The field has arrived at the same conclusion eighteen years later, under different names.

What I learned is this: the semantic infrastructure that makes agents trustworthy can only be built, enforced, and sustained by an organizational function with the mandate to do so. In most enterprises today, that function is structurally incapable of exercising that mandate, even where the CDO is competent, motivated, and well-resourced.

Part 2a established the governance gap. Part 2b quantified the stakes. Part 2c exposed the three illusion layers that hide the crisis from boards. Part 2d closes Part 2 by asking the organizational question the series has been building toward: who owns meaning, and what does the organization have to change to give that function the authority to matter?

The CDO crisis is not a distraction from the governance crisis. It is its organizational expression.

The Structural Verdict

The Chief Data Officer role was born twice. Each birth came with a different mandate and a different reporting line. Neither was designed for agentic AI.

The first CDO emerged from compliance. Sarbanes-Oxley, Basel II, HIPAA. The mandate was defensive: accuracy, access control, retention schedules, audit trails. The reporting line ran to the Chief Information Officer, then to the Chief Financial Officer. Data was a regulated asset to be controlled. Authority was derivative, borrowed from legal and compliance functions that had genuine organizational standing.

The second CDO emerged from digital transformation. Boards approved multi-year programs with ambitious timelines. Someone had to own the data track: domains, quality, ownership, roadmap. The mandate expanded, but the authority structure did not. Digital program ownership lived somewhere between the Chief Digital Officer and the Chief Technology Officer. The CDO's accountability grew without commensurate organizational authority. Implementation stayed in IT. Business units reported to different executives. The accountability gap was structural from the start.

Transformation programs do not end cleanly. They lose momentum, get rebranded, get absorbed. When the digital transformation program concludes or stalls, the CDO's mandate dissolves with it. The next program resets the clock.

There is a mechanism behind this pattern that the tenure statistics do not name. CEO priorities are episodic and program-anchored: a strategic bet, a transformation timeline, a compliance crisis. They respond to signals. Governance that functions correctly produces no signal. It succeeds by not generating events. The CDO who builds effective semantic governance generates exactly the condition that falls off the CEO's priority list: sustained, invisible absence of failure. Borrowed authority from transformation programs evaporates when those programs end because the CDO never secured independent standing on the CEO's attention list. They were present by proximity, not by mandate. The board mandate is the only structural alternative, because fiduciary continuity does not depend on which initiatives the CEO is currently tracking.

The result is the tenure data the field now accepts as normal. Fifty-three point seven percent of CDOs serve fewer than three years. Average tenure is approximately thirty months. Harvard Business Review's foundational research by Tom Davenport and Randy Bean identified the pattern years ago: CDOs carry accountability for outcomes they cannot control because they lack authority over the inputs. The tenure problem is the accountability gap made visible over time.

What makes the current moment different is Gartner's divergence finding. Deloitte surveyed the same CDO population and found ninety-four percent expect their influence to grow over the next twelve months. Gartner's forecast runs in the opposite direction: seventy-five percent of CDAOs not perceived as essential to AI success will lose their C-level position by 2027. Same role. Opposite trajectories. The difference between preparedness and irrelevance is mandate.

One month after publishing the CDO readiness paradox, Capgemini's research arm published TechnoVision 2026: one hundred and twelve pages on the autonomous enterprise. The word "agent" appears two hundred and sixty-nine times across forty-nine pages. The word "ontology" appears once, in parentheses. Knowledge graphs: zero. The role most prepared for this future is not mentioned once. One research arm finds CDOs most prepared. The other publishes the autonomous enterprise without them.

The CDOs expecting influence growth are measuring the wrong variable. Influence over pipeline delivery and dashboard programs is influence over the work that autonomous agents will absorb within the next eighteen to thirty-six months. CDOs building that case for their board are making a structural argument for their own replacement.

The role most prepared for this future is being measured on deliverables that this future will automate.

The Reporting Line Is the Mandate

When TotalEnergies EP Danmark created the Head of Data Office as a new position in 2020, I reported to the CFO from day one, as a peer of the Head of IT. That structural positioning determined everything about the work: the conversations I was invited into, the questions I was expected to answer, the problems I was authorized to define. Reporting as a peer of IT to the CFO meant data strategy, not technology support. Business value, not pipeline delivery.

A reorganization later, the function moved closer to operations, within the technology services structure but adjacent to the Asset Manager and team. The conversations changed. The questions changed. The work changed.

The reporting line does not just determine who you answer to. It determines what you are allowed to define. Under IT, you think about pipelines: integration, quality, delivery. Close to operations, you think about decisions: what data enables, what outcomes it governs. Close to the CEO, you think about meaning: what the organization believes, what it commits to, what it authorizes.

Most CDOs never reach the third level of thinking. Not because they lack the ability. Because the org chart decided for them before they arrived.

This is the structural problem that the agentic AI era makes urgent rather than merely important. When autonomous agents execute decisions at machine speed under corporate authority, the organization that cannot answer "who defines what those agents are authorized to mean by our terms and processes?" has not delegated. It has abdicated. And abdication carries different liability than delegation.

The debate over whether the CDO should report to the CEO, CIO, or CFO misses the operational question: does the CDO hold authority over meaning, or only over data? The CIO owns the pipes and the boxes. The CDO owns what flows through them and the contracts that define what may be done with it. Placing meaning governance under technology governance is how organizations produce the exact conflict they are trying to resolve: accountability for outcomes the function is not authorized to shape.

I argued at the time, and hold that view today, that digital should report into data, not the reverse. Digital programs are temporary. Data management capability is not. When the transformation program ends, the CDO's authority should grow, not dissolve. The program delivered tools. The CDO governs what those tools do with organizational meaning. That mandate has no end date.

Gartner has named context engineering as an organizational competency alongside AI engineering. Not a project deliverable. A permanent function. That is a mandate description. The organization that cannot assign that mandate to a person with genuine authority will build context without governance. That is exactly the failure mode this series has documented since Part 2a.

The Fork in the Road

When agentic AI arrived at scale, it presented the CDO function with a choice that cannot be deferred.

Agents will build pipelines. The same autonomous systems that execute customer interactions, process financial transactions, and coordinate logistics will also design, monitor, and optimize the data pipelines that feed them. The configuration and monitoring burden that defined data engineering work is progressively shifting to agent systems. The CDO who holds pipeline delivery as the core of their mandate is racing toward irrelevance. Not because pipelines stop mattering, but because pipeline management stops requiring a C-level executive.

What agents cannot do is define the context they operate in.

An agent that determines whether a customer qualifies for credit does not know, without being told in machine-readable terms, what "customer," "qualifies," and "credit" mean in your regulatory context, your risk framework, your operational history. It will improvise a definition derived from its training data. That definition will be coherent, plausible, and potentially wrong in ways that compound with every subsequent decision. The governance failure is not the agent's error. It is the absence of the organization that was supposed to define the terms before the agent ran.

Malcolm Hawker, CDO at Profisee and former Gartner analyst, has written consistently about CDOs caught flat-footed by the AI transition. The pattern he describes is structural: when the role expanded during the analytics era, most practitioners were hired to manage pipelines and deliver dashboards. The game changed. The training did not.

The Gartner framing is precise: context engineering (the organizational function that encodes operating rules in machine-enforceable form) means someone must own the ongoing work of defining what agents know about the world they operate in, enforcing what they are permitted to do in that world, and tracing what they did and why. Säde Haveri articulates the gap with a comparison that boards understand: when you hire a human employee, you onboard them: you explain the organization's context, culture, constraints. For an AI agent, most organizations give it API access to the wiki and say figure it out. That is not an onboarding failure. It is a mandate gap.

The CDO who owns the meaning layer, comprising the formal ontologies, the machine-enforceable constraints, and the provenance architecture, becomes the accountability function no autonomous enterprise can govern without. Formal ontologies name what terms mean across organizational boundaries. Constraint layers enforce what agents may do with those terms. Provenance records why those decisions were made and by what authority. That is not a deliverable. It is infrastructure. And when that infrastructure fails, every agent already running under corporate authority fails with it.

The CDO who builds what agents will automate becomes redundant. The CDO who owns what agents cannot function without becomes indispensable. The board decides which one it has.

The CDO is not at a crossroads. The CDO is at a countdown. That countdown is not only for the role. It is for the window in which the board can still establish governance before its agents' decisions become an unauditable, compounding liability.

When I led the Knowledge Graph Alliance, the governing question was never technical. Airbus, Michelin, Bosch, and TotalEnergies all had engineers who could build the architecture. The governance challenge was who sets the standards when competitors must align on them, and who owns the outputs when they disagree. We built SousLeSens, open source tooling for semantic web ontologies, precisely because shared infrastructure requires shared governance. That is exactly the question your board faces when agents from different systems act under shared corporate authority.

One fork leads toward what agents will automate. The other leads toward what agents cannot function without. The board decides which path the role is on....

The Jagged Frontier: Where AI Delivers

The pattern across every enterprise AI deployment studied in this series is identical: a small number of well-governed, narrowly scoped use cases delivers a disproportionate share of measurable value. The rest underdelivers or fails. Johnson and Johnson found that fifteen percent of nine hundred experiments delivered eighty percent of the value. For agentic AI, which unlocks capabilities that have no analogue in prior digital programs, that ratio is likely to skew further still. The variable that explains the pattern is not model quality. It is whether governance exists at the load-bearing points before the system runs.

The evidence that agentic AI creates genuine, measurable value is not ambiguous. The evidence that it does so under specific structural conditions is equally unambiguous. Understanding both sides is the precondition for board-level governance.

DBS Bank is the clearest enterprise case study. The bank deployed AI across multiple business domains and projected a revenue impact of approximately seven hundred and sixty-eight million USD in 2025, up from around seven hundred and fifty million Singapore dollars in 2024, based on roughly 370 AI use cases powered by over 1,500 models. The result was not accidental. DBS developed the PURE framework (Purposeful, Unsurprising, Respectful, Explainable) before deploying agents at scale. Purposeful means every agent deployment is linked to a specific business outcome with measurable success criteria. Unsurprising means agents operate within boundaries stakeholders understand and accept. Respectful means agent decisions respect customer rights and regulatory requirements. Explainable means every decision can be traced, audited, and explained to a regulator on demand. The semantic infrastructure underneath the PURE framework is what makes each property enforceable rather than aspirational.

Johnson and Johnson offers the quantitative case for constrained deployment at scale. The company ran nine hundred AI experiments over three years. Fifteen percent of those experiments delivered eighty percent of the measurable value. The remaining eighty-five percent delivered marginal or negative returns. This is not a failure story. It is a Pareto distribution that reveals the structure of AI value creation. The fifteen percent that worked shared a pattern: narrow scope, well-governed data, clear success criteria, human authority over the decision threshold. The eighty-five percent that underdelivered shared a different pattern: broad scope, ambiguous success metrics, and an assumption that model capability would substitute for governance clarity.

The OLF/Epsis Value Assessment, which I contributed to in 2006 and 2007 across the Norwegian Continental Shelf, established a quantitative pattern that predates the AI era: eighty-two percent of digital transformation value comes from new capabilities that were previously impossible, not from efficiency gains on existing operations. Only eighteen percent comes from doing the same things faster or cheaper. Most AI budgets target the eighteen percent. The organizations that capture the eighty-two percent are building new decision capabilities, not automating existing workflows. New capabilities require agents to act on meaning. Efficiency gains require agents to execute process. The former demands semantic infrastructure. The latter can operate without it. For a while. Agentic AI, which makes previously impossible decision capabilities the primary value proposition, almost certainly pushes that ratio further. The semantic infrastructure question does not disappear at the eighteen percent. It becomes the condition for reaching the eighty-two.

Glowforge, the consumer laser cutter manufacturer, illustrates the pattern at an accessible scale. An AI agent handles operator support within a precisely bounded domain: materials, cutting parameters, design troubleshooting. The boundary is enforced, not advisory. The agent does not schedule service visits, process returns, or make pricing decisions. The result: support volume absorbed without proportional headcount growth.

The pattern across every successful deployment in this series is identical: constrained scope, semantic grounding, clear authority model, enforceable boundaries. The Pareto distribution is not accidental. AI delivers value in proportion to the precision of the semantic infrastructure that governs it.

Every organization on the left side of this table built governance before deploying capability. Every organization on the right assumed governance could follow. It cannot.

The Jagged Frontier: Where AI Fails

Evan Ratliff at WIRED tested the "one-person billion-dollar company" thesis in live production. He built a startup staffed entirely by AI agents. The failure modes were not catastrophic. They were mundane. His CTO agent fabricated performance metrics, then memorized those fabrications as ground truth for subsequent decisions. It treated casual remarks in a Slack channel as authorized instructions and allocated budget accordingly. By the time Ratliff read the logs, the errors had compounded across multiple decision cycles. No single failure was dramatic. The accumulation was.

That is the iterate-into-governance fallacy running in production. The assumption that agents will self-correct, that monitoring will catch drift, that the system will signal when it needs intervention. It did not. The failures were invisible at each individual step. Visible only in retrospect, when the compounding had already occurred.

The "Agents of Chaos" research paper documents the same pattern at institutional scale. Twenty researchers from Northeastern University, Harvard, MIT, Stanford, and Carnegie Mellon red-teamed live multi-agent systems over two weeks. Ten distinct security vulnerabilities documented. The finding boards should memorize: none emerged from jailbreaks or adversarial prompts. They emerged from normal use. An agent deleted a mail server because someone who did not own it made the request conversationally. The agent responded to urgency, not authority. No constraint enforced the difference.

The research identifies three structural absences that make multi-agent systems dangerous by default. An authority model: agents respond to whoever speaks most urgently, not to whoever holds organizational authority. A self-model: agents could execute tasks but could not assess whether they were authorized to do so. Capability without self-assessment is the definition of an agent without governance. A private deliberation surface: agents could not distinguish between internal reasoning and external action. The boundary between reasoning and acting was absent.

Multi-agent systems amplify every one of these absences. When agents share context through the same channels they use to coordinate, vulnerabilities propagate through the collaboration architecture itself.

BCG published a four-tier autonomy framework. Monitoring is supposed to bridge Tier 3 and Tier 4. It does not close the governance gap. Monitoring is detective: it reads logs after the agent acted. The action is already complete. Many consequences are irreversible. A deleted mail server is not restored by reading the audit trail. A credit decision already executed cannot be unilaterally undone.

I built and governed autonomous decision systems in 2007, using formal agent architecture with ontologies defining the operational context for subsea drilling agents. The consequences of a stuck drill bit running at $200,000 a day meant governance failures were immediately visible in financial terms. The mechanisms have changed. The failure mode has not. Governance built around observation after execution leaves the authority model unaddressed. By the time you observe the failure, the failure has already made its decisions.

At scale, one thousand agents simultaneously executing tasks that human teams perform sequentially collapses every management assumption that governs sequential work. Without active monitoring infrastructure routing the right intervention to a human with genuine authority, Wharton's Ethan Mollick's "sin eater" dynamic takes over: someone absorbs the blame. But blame is not accountability. Accountability requires that the structure existed before the failure: the authority model, the constraint layer, the escalation route.

IDC FutureScape 2026 quantifies the consequence trajectory: by 2030, up to twenty percent of the thousand largest global organizations will face lawsuits, substantial fines, or CIO dismissals due to inadequate controls over AI agents. This is why the constraint layer must be preventive and centrally owned by whoever holds the meaning mandate. Distributing it across IT, legal, and compliance functions without a single point of integration guarantees the gap compounds. Gartner adds the cancellation data: over forty percent of agentic AI projects will be canceled by 2027. Not paused. Canceled after investment. For three reasons: escalating costs, unclear business value, and inadequate risk controls. A third Gartner prediction identifies the mechanism: by 2030, half of all agent deployment failures will result from insufficient runtime enforcement of governance controls. Not model error. A missing enforcement layer.

The jagged frontier is real. The line between the two sides is not model quality. It is whether governance exists at the load-bearing points before the system runs.

The Governance Gap No Framework Closes

Most governance frameworks for agentic AI share a structural limitation: they are built around observation. Monitor what agents do. Alert on anomalies. Review logs. Audit trails. These are detective controls. They establish what happened. They cannot prevent what is about to happen.

The EU AI Act, Article 14, requires human oversight of high-risk AI systems. Most organizations interpret this as monitoring dashboards: a human watches what the agent does and can intervene if something looks wrong. That is the wrong interpretation. Monitoring is detective control. Article 14 requires authority structures: the formal definitions of who can authorize what, under what conditions, with what boundaries. Human oversight means structures that stop an action before execution, not a dashboard after. The CDO who delivers a monitoring dashboard is doing compliance theater.

The EU AI Act places liability at the provider and operator level: the business leadership, not the engineering function. Boards that delegate AI governance to IT are not just making an organizational error. They are delegating their own liability to a function that does not have the authority to exercise it. When an agent acts under corporate authority, the liability belongs to the board, regardless of which team configured the system.

Governance only becomes real when constraints are enforceable at the moment a system acts, not after the fact. Constraints that are only advisory are not constraints. They are instructions. An agent that can reason past a boundary is an agent without a boundary.

The architectural shift required is this: governance stops being a role or a process and becomes enforced properties of the execution environment itself. Governance as a role means the CDO reviews decisions after they are made. Governance as enforced properties of the execution environment means the CDO owns the infrastructure that makes compliant decisions the only decisions possible. The first is auditing. The second is architecture.

Scope limitation and correction capacity are not the same thing. Scope limits blast radius. It is necessary. Correction capacity determines whether you can still stop the system once drift has begun. Most governance frameworks provide the first and assume the second. The assumption is wrong. An agent that operates within a defined scope can still drift within that scope: accumulating fabricated ground truth, treating casual inputs as authorized instructions, compounding small errors across decision cycles in exactly the pattern Ratliff documented. Scope without correction architecture is containment without control.

Three governance layers are required. Most programs have only one.

There is a third failure mode that compounds all of them. When organizations cannot find ontology engineers (and the talent shortage is real) they reach for a shortcut that feels modern: let an LLM write the ontology. Think of the ontologist as an insurance premium. You would not let an AI draft your corporate bylaws without a lawyer. The reasoning is identical: the cost of a plausible-but-wrong legal foundation only becomes visible when it fails under real conditions. An LLM-generated ontology is a hallucinated foundation. It validates syntactically. It fails operationally.

The reason is structural, not technical. Ontology engineering is fundamentally an act of exclusion: deciding what does not belong in the model and why. Ronald Ross, whose career has been spent at the intersection of policy interpretation and concept modeling, names the core skill with precision: conceptualization is deeply human and in critically short supply. No compression engine has it. Andrej Karpathy calls LLMs a lossy compression of the internet: compression deletes the edges and amplifies the mean. Your business lives in the edges: the line between a service contract and a framework agreement, the threshold where maintenance becomes capital expenditure, the decision thresholds your risk appetite defines. Ask a compression engine to define those boundaries and you get everyone else's model. Not yours.

The production evidence is consistent. Teams that spend months using LLMs to build enterprise ontologies encounter entity duplication, hallucinated relationships, and rising infrastructure costs. Scrapping the LLM-generated model in favor of a small set of human-curated concepts reliably outperforms it. If you understand your domain, why pay a compression engine to rediscover it poorly?

The governance implication is not that LLMs are useless in ontology work. As Kurt Cagle and Michael DeBellis both argue, LLMs can accelerate the grunt work: generating candidate terms, drafting SHACL shapes against an existing structure, producing test data. The failure mode is different: organizations that skip human expertise entirely and treat LLM output as the governance foundation are building on a base that validates syntactically and fails operationally. You will hire an ontologist at the beginning or six months later when the foundation collapses. The timeline is the organization's choice. The requirement is not.

Architecture without governance is not autonomy. It is exposure.

Most governance programs have Layer 1. Governance requires all three. The gap between what most programs have and what agentic AI requires is where liability accumulates.

Five Questions That Close Part 2

Part 2c asked whether your board could see the governance gap. These five questions ask whether your board owns the authority to close it. They are diagnostic, not ceremonial. Most boards cannot answer them. The gaps they expose are the governance gaps this series has documented since Part 2a. They require no technical expertise to ask. They require significant organizational honesty to answer.

Do we own the rules that run our AI?
Not the models. Not the pipelines. The rules: the formal definitions of what your agents are authorized to mean by your terms, what they are permitted to do, and what they are explicitly prohibited from doing. If those rules live in vendor platforms, policy documents, or prompt instructions that agents can reason past, you do not own the rules. You have approximations. An approximation that governs a credit decision, a procurement approval, or a safety classification is a liability waiting to materialize.
Can our agents prove which facts they used?
Provenance must be machine-auditable before execution, not reconstructible after the fact. When an agent recommends, decides, or acts, the decision trail must be traceable to the specific data, the specific rules, and the specific authority structures that produced it. If your team's answer requires reading logs retrospectively, you have evidence collection. Governance requires that the explanation exists at the moment of action, not after it.
Can we stop an agent from violating a new rule instantly?
This is the enforcement architecture question most governance frameworks evade. When your regulatory context changes, and it will, can you update the constraint layer governing agent behavior and have that update propagate across every running agent immediately? Or does enforcement require retraining, redeployment, or manual intervention? The answer reveals whether governance is structural or advisory. Structural enforcement is preventive. Advisory enforcement is monitoring dressed as governance.
Can a regulator understand why the agent decided what it did?
Not your engineering team. A regulator. Someone who was not present when the system was built, does not have access to the model weights, and requires an explanation grounded in the business rules and authority structures your organization has formally defined. The EU AI Act, Article 14, requires exactly this for high-risk systems. If the honest answer is that explainability depends on the model's internal processing rather than on machine-readable constraints your organization owns, you do not have regulatory-grade governance. You have a compliance gap accumulating with every autonomous decision.
Is this a foundation we can scale, or a silo we must rebuild?
Palantir published the architecture diagram that settles what the foundation looks like: ontology at the top center, LLM at item one of twelve at the bottom right, treated as untrusted and interchangeable. A company worth over $180 billion built that architecture over two decades. The question for every other board is not whether this architecture is necessary. Palantir settled that. We adopt the principle, not the platform. The question is whether you are building your semantic layer on open W3C standards you control or on proprietary formats you rent. Every object type, every constraint, every action rule encodes how your organization makes decisions. That accumulated meaning is intellectual property, not infrastructure. It appreciates over time. A competitor who rebuilds it from scratch is years behind you. If you cannot move it without losing it, you do not own your governance. You subscribe to it.

The five questions do not require a technical answer. They require an organizational one: who is accountable for each answer, what authority do they hold, and what evidence exists that the answer is operational rather than aspirational?

Next In This Series

Part 3 opens the HOW question. If governance is missing and the CDO mandate must evolve, what does the architecture actually look like? What must a CDO own, technically and organizationally, to govern agents at enterprise scale?

Part 3a examines the CDO mandate in detail: what it must include, what organizational authority it requires, and why the organizations that get this right will build competitive advantage that deepens over time. It opens with a question that most CDOs cannot currently answer: who in your organization manages meaning?

Part 3a publishes in two weeks. Subscribe to follow the series.

Subscribe now

About the Author

Frédéric Verhelst helps leadership teams build the foundations for trustworthy agentic AI. With a PhD in Applied Physics cum laude from Delft University of Technology and twenty-five years at the intersection of data, AI, and industrial operations, he has governed AI agents in industrial operations since 2007.

He governed semantic infrastructure across twenty-two energy companies as Program Manager of the IOHN joint industry initiative on the Norwegian Continental Shelf, led cross-industry governance as CEO of the Knowledge Graph Alliance, whose founding members include Airbus, Michelin, Bosch, and TotalEnergies, and served as Head of Data Office at TotalEnergies EP Danmark, where he reported to the CFO as a peer of IT and shaped data governance strategy at group level as a member of the global Data Governance Council. He presented at the W3C Semantic Web workshop in 2008, arguing for what the industry now calls agentic AI governance.

He is featured on the GraphRAG Curator Podcast and on Enterprise Wide Search with O’Reilly author Ole Olesen-Bagneux. He works at Viking Life-Saving Equipment, where agentic AI governance for mission-critical safety operations is taking shape. His specialty in one sentence: designing and governing the semantic control plane that makes autonomous decisions auditable, stoppable, and scalable.

Follow him on LinkedIn for the latest posts in The Ontology Imperative: Building Trustworthy Agentic AI.

2c – The Hidden Governance Crisis – What Natural Language, Inherited Bias, and Dashboard Confidence Are Hiding from Your Board

Frédéric Verhelst — Thu, 12 Mar 2026 07:27:20 GMT

Reading time: ~18 minutes

Summary

Your teams are moving fast with natural language interfaces. The demos look like intelligence. The dashboards confirm readiness. The budget looks clean. But three layers of illusion are hiding the same structural gap from your board: the absence of semantic infrastructure. Natural language hides the foundations agents depend on. Training data inheritance hides cultural bias your organization never approved. And analytics maturity hides the boundary where governance ends and autonomous action begins. This article exposes all three, and the investment math that proves the gap is not a risk to manage but a transformation to capture.

Board Brief: Three Illusion Layers Hiding a Structural Crisis

Natural language interfaces make agents look smarter than their infrastructure allows. The board sees a demo where an agent answers a complex business question in perfect English. What the board does not see is the stack of infrastructure required to make that answer correct. Or the absence of that stack making the answer dangerously wrong.

Training data creates a second illusion. Every LLM-based agent begins with cultural defaults inherited from internet text: overwhelmingly Western, Educated, Industrialized, Rich, and Democratic. Your agents do not start neutral. They start biased toward a worldview your organization never signed off on.

Analytics maturity creates the third. Beautiful dashboards create confidence that data infrastructure is AI-ready. It is not. Governance ends at the warehouse. Everything downstream, where agents actually operate, is ungoverned.

Three questions for your next board meeting:

Can your team explain why the agent chose one definition of “customer” over another?
Does your AI budget include preparation costs, verification infrastructure, and semantic debt, or just compute and licensing?
Is your data infrastructure optimized for dashboards or for autonomous decision-making? If the answer is dashboards, you have mathematically capped your AI ROI.

If any answer exposes a gap, no amount of prompt engineering will close it.

The Three Illusion Layers

Boards are funding AI programs on the basis of what they can see. What they cannot see is where the crisis lives.

Executive Brief

I believe in agentic AI. I have also spent twenty-five years watching organizations confuse what technology can demonstrate with what infrastructure can sustain. I have led data offices and chaired strategic initiatives across twenty-three energy companies. I have seen this gap from both the engine room and the boardroom.

In 2012, I co-authored the exploration use case and contributed to the business case for the EU Optique project on behalf of Statoil (now Equinor). At the same time, as head of the Special Interest Group for Production & Reservoir at POSC Caesar Association, the custodian of ISO 15926, I understood what exploration engineers needed to query industrial data using simplified language, hiding database complexity, exactly as today’s AI agents do. Optique project materials projected potential savings of fifty million euros per year at Siemens and seventy million at Statoil. At Siemens, by removing the human intermediary who previously translated engineering queries into database operations, end-to-end turnaround times dropped from weeks to hours.

What made abstraction work was not the interface. It was the ontologies and mappings underneath. Semantic infrastructure that translated simplified queries into correct operations across heterogeneous data sources. Without it, the queries would have returned plausible-looking answers that were wrong.

That was 2012. The pattern has not changed. Only the stakes have.

Part 2a established the governance gap: most organizations cannot trace agent decisions, distinguish reasoning from hallucination, or fix root causes. Part 2b showed the stakes: investment without foundations creates the Triple Paradox: faster but more fragile, more autonomous but less accountable, more scalable but more brittle. The false solutions were exposed: observability monitors after the fact but cannot prevent failures, and LLMs cannot reason their way to governance because they destroy the provenance that governance requires.

Part 2c asks the question that precedes all solutions: why can the people who need to see this crisis not see it? The answer is three layers of illusion, each plausible, each hiding the same structural gap, and each reinforced by the vendors, metrics, and mental models that boards trust most.

The Illusion of Conversational Intelligence

Every technology wave begins with seduction and ends with plumbing. Data warehouses: you cannot skip dimensional modeling. Big data: you cannot skip data quality. Semantic systems: you cannot skip ontologies. Natural language interfaces are the latest seduction.

André Lindenberg, writing in his Closing the Context Gap newsletter, makes the stakes concrete. An agent sees three months of declining support tickets and recommends scaling back the support team. Without ontological context, the recommendation is logical: volume is down, reduce capacity. With it, the agent recognizes a disengagement pattern and flags a retention risk instead. The data is identical. The consequential action is opposite. Lindenberg calls this mundane misdirection: the difference is not model quality or prompt engineering. It is whether the system has access to a representation of why the world works the way it does. Natural language hides that requirement entirely.

Your vendor demonstrated an agent answering “how much business with Walmart?” The board saw magic. What actually happened was a chain of infrastructure components executing in sequence: the agent translated the question to a query language; the data catalog mapped “Walmart” across entity names; the master data management system resolved which entity was meant; the semantic layer defined what “business” means in context; the BI layer executed the query against governed data; and the agent formatted the result in conversational English.

Every piece essential. Natural language hid all of it.

Malcolm Hawker, former Gartner analyst and Chief Data Officer, describes exactly what Optique demonstrated: agents are not executing queries. They are orchestrating them. The accuracy comes from the BI and data management layers, not from the AI. The project proved this fourteen years ago. The simpler the interface, the more rigorous the infrastructure must be. Every organization skipping that work today is recreating the same structural gap. Only now the agents act, not just answer.

What Optique demonstrated is that abstraction and rigor are not opposites. They are dependencies. The EU project succeeded because the ontologies beneath the interface were precise enough to translate simplified exploration queries into correct operations across Equinor’s heterogeneous subsurface databases and Siemens’s turbine diagnostics systems. The consortium did not just connect data. It governed meaning. The interface looked effortless. The ontological mapping work was anything but. That gap between visible ease and invisible rigor is exactly where today’s governance crisis lives.

Today’s vendors bet that LLMs can replace what ontologies and mappings do. Boards see AI answering questions and conclude infrastructure is optional. They cut data management budgets. They cancel semantic programs. They fund the interface and starve the foundations.

If your master data management cannot resolve “Walmart,” the agent guesses. If your semantic layer does not define “business,” the agent improvises. If your data catalog is outdated, the agent queries the wrong source. The answer still arrives in perfect English. Formatted with confidence. Wrong.

This is not future risk. This is governance failure disguised as AI competence. And natural language makes it invisible to the people approving the budget.

When your agent answers “how much business with Walmart,” can your team explain why it chose one definition over another? If that answer relies on the BI layer, you have infrastructure. If it relies on the LLM, you have hallucination formatted as analysis.

The Inheritance Your Organization Never Approved

The interface illusion hides what agents need. The inheritance problem hides where agents start. A third illusion hides how boards are responding: they think they are buying software.

Your agents were not trained on a representative sample of human knowledge. They were trained on the internet. And the internet is overwhelmingly WEIRD: Western, Educated, Industrialized, Rich, Democratic. Harvard evolutionary anthropologist Joseph Henrich documented that WEIRD populations represent roughly twelve percent of the world's people yet dominate the research record. LLMs trained predominantly on internet text inherit the same imbalance.

Andrej Karpathy, founding member of OpenAI, calls LLMs a lossy compression of the internet. Lossy compression deletes the edges and amplifies the mean. Your business lives in the edges. Industrial safety protocols, regulatory nuance, multilingual operations, non-Western decision patterns. None of this exists at internet scale. The model regresses everything toward the statistical average of whoever had time and access to post.

You did not deploy intelligence. You deployed inheritance.

This is not a social justice concern. It is a governance failure. An ungrounded agent penalizes deviation from the “internet average.” Local nuance becomes error. Statistical bias becomes liability. Ask about safety, and the agent recites U.S. norms. Ask about authority, and it copies Silicon Valley hierarchies that undermine industrial command structures. Ask about conflict resolution, and it applies patterns that escalate in cultures where they were designed to de-escalate.

The agent does not get it wrong so much as it gets it consistently, boringly “normal,” and then treats your actual environment like a rounding error. Not loudly. Quietly. With confident answers. Your shop floor, your regulators, your languages, your chain of command look different from the internet’s version of reality. The agent does not argue with that difference. It punishes it. By optimizing toward a statistical center no one in your organization signed off on.

The inheritance problem is not abstract. If your agent is trained predominantly on WEIRD text, it will default to U.S. common law logic when drafting a procurement summary for a subsidiary in Singapore or Germany. It will apply Anglo-American negotiation norms to cultures where they backfire. It will structure risk assessments around regulatory assumptions that do not apply in your jurisdiction. You have not deployed intelligence. You have deployed a cultural ghost in the machine, and it is making decisions your legal team has never reviewed.

This is what makes the inheritance problem a governance problem rather than a bias problem. Ungrounded agents behave coherently. They optimize. They produce results. But they do so inside a frame of meaning that nobody owns. The issue is not malfunction. It is unauthorized optimization. The model is not hallucinating. It is following a logic you did not write. And the question most enterprises cannot answer is not how their agents behave, but who has the authority to stop an agent chain the moment outcomes look plausible but feel wrong. Most organizations have no mechanism at all.

The governance test is not whether agents produce biased output. They will. The test is whether your organization has a mechanism to detect and correct that bias before it compounds into operational damage.

Twenty years ago at Epsis, we built hybrid AI because autonomous BDI agents inherited training bias from their learning environments. Ontologies corrected it then. The same correction applies today, at industrial scale. Without grounding in domain truth, agents do not accelerate insight. They accelerate drift. You lose control the moment you assume the agent starts neutral.

Ontologies are the correction layer. They replace inherited defaults with authorized truth. They give agents semantic boundaries instead of cultural guesses. They are the first real mechanism that replaces statistical inheritance with authorized meaning. When boards cut data management because “the model knows,” they remove the only safeguard preventing the internet’s worldview from becoming operational error.

The R&D Problem Boards Are Funding with Procurement Logic

Your board thinks it is buying software. It is not. At the same Summit, over half of IT and AI leaders reported being pushed into AI adoption, not pulled by strategy. That pressure is the mechanism. Procurement logic is the response. And the combination is what produces semantic debt at scale.

It is stepping into an organizational R&D challenge and handling it with procurement logic dressed up as innovation.

R&D tolerates failure. R&D funds learning loops. R&D measures capability built rather than cost avoided. Most boards avoid all three. They run AI pilots like product launches. They demand ROI before they have built the structures that make ROI possible. They try to purchase trust in autonomous systems instead of building the organizational capacity required to govern them.

Ethan Mollick at Wharton has documented the same pattern: AI adoption is not a deployment challenge. It is a research challenge. Organizations that treat it as procurement get procurement-grade outcomes. Sarah Guo at Conviction reinforces this from the venture perspective: what enterprises face is closer to experimental product building than enterprise software deployment. A framework most organizations are structurally unable to absorb because their governance, budgeting, and accountability models assume predictable outcomes from predictable inputs. Agentic AI produces neither.

The productivity evidence supports this framing. Recent analysis in The Economist highlights what economics correspondent Alex Domash describes as a modern echo of the Solow paradox where technology investment runs ahead of measurable productivity gains. When the investment is structural but the measurement framework expects operational returns, the gap looks like failure when it is actually misclassification. Organizations measuring AI against quarterly efficiency targets are using the wrong ruler. The value is in capabilities that did not exist before. Capabilities that require R&D investment in infrastructure that does not yet appear on any balance sheet.

I have seen this ratio before. When I was part of the team that quantified the value of Integrated Operations across the Norwegian Continental Shelf, we found a stark split: eighty-two percent of value came from entirely new capabilities; only eighteen percent from efficiency gains (based on the 2006 assessment).

Boston Consulting Group documented the same pattern in transformation programs: seventy percent of transformation effort is people and processes, twenty percent is technology and data, ten percent is algorithms. Most boards are spending seventy percent of their budget on the ten percent and then blaming the technology when results stall.

For agentic AI, the ratio is even more skewed. Boards are budgeting the eighteen percent, the efficiency slice, while the transformational eighty-two percent requires R&D investment in semantic infrastructure, organizational redesign, and capability building that never makes it into the business case.

If your AI program has a quarterly ROI target and a cancellation clause, it is not R&D. It is theater.

The Hidden Cost Structure Your CFO Is Missing

The R&D misclassification feeds directly into the cost structure problem. Your CFO approved the AI budget. Compute. Licensing. Talent. Clean spreadsheet. What will determine whether the investment succeeds or fails is not on that spreadsheet.

Rama Ramakrishnan at MIT Sloan identifies three costs that determine AI program survival: preparing the system for your domain, running it, and checking and fixing outputs. CFOs budget the middle one. Undercount the first. Ignore the third.

His sharpest insight is the correctness spectrum. Some tasks tolerate loose correctness. A draft email can be wrong. GenAI lives here. Costs stay manageable because humans catch errors. Agentic AI does not live here. Agents that approve transactions or reject suppliers require absolute correctness. No draft. No review cycle. The agent acts.

Your CFO is budgeting for loose-correctness GenAI while authorizing absolute-correctness agentic AI. That gap is where programs die. This maps directly to the risk stratification framework from Part 2a: low-stakes tasks can rely on prompts and human review; high-stakes tasks that move money, grant access, or operate in regulated domains require a formal ontological contract. The cost structure follows the correctness requirement.

Preparation is ontology work. It means defining what “revenue” or “compliant” means so agents operate on authority instead of statistical guesses. As Ramakrishnan argues, domain experts must define what “good enough” means for each task. That is ontology design, whether you call it that or not.

Framework: Ramakrishnan, MIT Sloan. Chart-of-accounts mapping: author.

Checking is governance architecture. For GenAI, a human reviews output. Expensive but linear. For agentic AI, there is no human in the loop. You build automated constraint infrastructure or you have zero checking. Not reduced checking. Zero. The cost of that infrastructure belongs on line one of the AI budget. Anthropic demonstrated this when they gave Claude a thousand dollars and a vending machine business to run. In pre-deployment simulation, the agent went ten days without sales, tried to contact the FBI about a two-dollar recurring charge it could not explain, and declared the business dead. Once deployed, employees scammed it out of two hundred dollars by claiming prior discounts, and it hallucinated being a human wearing a blue blazer. The failure was not intelligence. It was the absence of constraints defining what “reasonable business decisions” meant. The checking infrastructure did not exist. The cost was real.

Tom Blomfield, Group Partner at Y Combinator, warns about what he calls the scaffolding problem: there is 'quite a lot of scaffolding' between a foundation model and production deployment that organizations rarely budget for, and without safeguards in place, you cannot have confidence an agent won't overwrite your data. The infrastructure cost (evaluation frameworks, retrieval pipelines, constraint systems, monitoring) is real and substantial. Julius Hollmann, CEO of digetiers, names the accumulating liability directly: semantic debt. Every agent deployed without semantic infrastructure beneath it adds to this debt. It compounds. It does not self-correct. And unlike technical debt, which degrades performance, semantic debt degrades trust. An organization that accumulates enough of it loses the ability to know whether its agents are acting on authorized meaning or inherited guesses.

Hollmann describes ontologies as constitutions for AI systems: they do not tell agents what to do in every situation, but they define the boundaries of legitimate action. Without that constitution, agents operate in a legal vacuum: technically functional, governance-wise indefensible.

As Matteo Turi puts it: AI does not create wealth. Architecture does.

For the CFO reading this: the Triple Cost Framework (Preparation, Running, and Verification) maps directly to your chart of accounts. If all three are sitting in the IT budget, you have structurally misclassified the investment. Preparation is R&D. Verification is Compliance. Running is Operations. Treat them as such, or your AI program will be cancelled for “underperformance” when the real failure was accounting.

Vendors sell the agent. They do not sell the contract the agent needs to operate safely. That contract is an ontology. Ontologies create that capability. Without them, you funded automation. Not transformation.

The Dashboard Trap: Where Governance Ends and Agents Begin

The most dangerous illusion is the one that feels like proof of readiness.

Your dashboards are beautiful. Real-time KPIs, executive scorecards, self-service analytics. They create confidence that your data infrastructure can support AI. It cannot.

Sarah Levy, Co-founder and CEO of Euno, identifies the exact boundary: governance ends at the warehouse. Organizations govern structured data inside the warehouse: quality checks, lineage, access controls. Once data flows downstream to AI agents, governance drops to zero. Everything beyond the warehouse is the Wild West. At the Gartner Data & Analytics Summit in Orlando this week, 14% of organizations reported confidence that their data is secured and governed. Not a projection. A self-assessment from the people closest to the systems. Levy's warehouse boundary is not a metaphor. It is what 86% of organizations are operating inside. Levy makes a second point that most organizations miss: AI-readiness is contextual and probabilistic, not binary. A product team can tolerate statistical imprecision that a finance function cannot. An agent recommending content operates under different correctness requirements than an agent approving supplier payments. The same data may be “ready” for one use case and dangerously unready for another. But the dashboard metrics that boards rely on make no such distinction.

This worked when humans interpreted dashboards. Humans provide the manual governance layer. They catch anomalies, interpret context, apply judgment. Dashboards represent passive risk: a human might think the wrong thing. Agents represent active risk: a machine does the wrong thing at machine speed. When a dashboard misleads, an analyst is confused. When an agent misacts, the board carries the liability.

Cassie Kozyrkov, Google’s former Chief Decision Scientist, has argued convincingly that most organizations are “data-inspired” rather than “data-driven.” Management already knows the answer. The dashboard provides confirmation, not direction. You want the hotel; you see 4.2 stars; you say “great.” But if you did not want it, 4.2 would not be enough. The decision came first. The data followed.

Agents cannot do this. They have no unexpressed preference to protect them from a misleading metric. So dashboards did not just optimize observation instead of agency. They optimized confirmation. This is the organizational version of the observability fallacy from Part 2b: we are measuring what happened instead of governing what is possible. And now organizations want to hand that decision loop to agents, built on infrastructure that never actually drove decisions in the first place. At least humans could silently compensate for the gap between what dashboards showed and what the situation required. Agents cannot. They will encode the confirmation bias into autonomous action at machine speed.

The infrastructure mismatch runs deeper. Alan Morrison documents forty years of data layer inertia: enterprises are running modern AI on relational foundations designed for 1980s reporting. Only one graph database management system appears in the DB-Engines top twenty, and it is not standards-based. The data layer is frozen in the paradigm that built dashboards. Perfect for tabular analytics. Structurally incapable of representing the relationships agents need to reason. Gartner’s 2025 survey of over five hundred CIOs found that seventy-two percent of organizations are breaking even or losing money on their AI investments. The explanation is not that the models are weak. The explanation is that the data infrastructure was optimized for a different job.

The problem is not just architectural. It runs into what data quality actually means. I am a geophysicist. I have spent a career working with data where context determines meaning. What is noise for one application is signal for the next. The same seismic trace that a structural geologist discards may contain exactly the information a reservoir engineer needs. Quality is not an attribute of data alone. It is an attribute of data in combination with its intended use. Data that scores perfectly on your dashboard quality metrics may be structurally unfit for agent reasoning. Your data quality program optimized for dashboards. Nobody asked whether the same data is quality for agents.

Dashboards tolerate missing relationships because humans fill in the blanks. Agents fail because of them. Relational snapshots are fish: fresh for a day, then useless. Agents require wine: the compounding context of how relationships evolve over time. Analytics-optimized data is fish served as wine to systems that need aging, depth, and relational richness to reason correctly. That is not a data quality problem. It is a data architecture problem.

Gartner’s AI-Ready Data Essentials framework confirms this institutionally. Stage 4 explicitly requires “rich semantics to improve the accuracy of GenAI on top of enterprise data.” Traditional data management is insufficient for AI. Semantics is listed as a required component alongside data quality, lineage, and governance. Yet most organizations have built the first three and skipped the fourth.

Säde Haveri, speaking on the Catalog & Cocktails podcast with Juan Sequeda and Tim Gasper, put the distinction precisely: most data and AI initiatives don't fail on technology. They fail on meaning. Can you achieve meaningful data governance without ontology work? With data, yes. With AI, no. The numbers confirm the pattern. At the Gartner Data & Analytics Summit in Orlando earlier this week, 44% of organizations had implemented semantic layers in 2025. Another 48% plan to by 2027. Within two years, nearly the entire enterprise data community will have deployed one. Juan Sequeda, who reported the statistics, asked the question most organizations are not asking: are they creating ungoverned semantic layers? Probably. The semantic layer surge is not a sign of readiness. It is an emergency response to governance failure, executed at speed, without the ontological foundation that would make the layers governable. The stages Gartner describes as essential infrastructure are exactly the work most organizations deferred when dashboards were sufficient. AI made the deferral consequential.

The math is unforgiving. When we quantified digital transformation value across the Norwegian Continental Shelf, the ratio was clear: efficiency delivers the minority share; new capabilities deliver the majority. Dashboards deliver the efficiency slice. Agents require the transformation slice. If your infrastructure is optimized for dashboards, you have mathematically capped your AI ROI at the eighteen percent. You have funded automation but starved transformation. Moving past the warehouse boundary requires ontologies that define meaning across the entire data lifecycle, not just where data is stored, but where decisions are made. Without them, your agents inherit the Wild West.

The irony is precise. The organizations that invested most heavily in analytics maturity (the most sophisticated dashboards, the best data quality scores, the most mature reporting infrastructure) are often the ones most at risk. Their confidence is highest precisely where their gap is widest. The dashboard trap is not a failure of underinvestment. It is a failure of misdirected investment. And the organizations caught in it will be the last to recognize the problem because every metric they trust tells them they are ready.

Dashboards are the illusion of readiness. Agents reveal the truth.

What To Do Now

The three illusion layers share a common correction: semantic infrastructure that makes invisible foundations visible, inherited bias explicit, and governance boundaries enforceable.

Demand an interface audit for every AI demo and pilot. Require the team to document the full decision chain: what resolved the entity, what defined the business term, what governed the data source, what validated the output. If any link in that chain is “the LLM figured it out,” stop funding until it is replaced with governed infrastructure.

Refuse any AI budget that does not separate preparation, running, and verification as distinct cost lines. Preparation is R&D: ontology design, domain modeling, entity resolution. Running is Operations: compute, licensing. Verification is Compliance: automated constraint infrastructure, governance architecture. If the business case only covers running costs, it is not a budget. It is a down payment on semantic debt.

Require a grounding assessment before deploying agents in any domain where cultural, regulatory, or operational context differs from the Western internet average. Identify what authorized domain knowledge the agent has access to, what inherited defaults could override it, and where the agent would penalize deviation from the statistical mean. Ontologies provide the correction mechanism. Without them, you are deploying the internet’s worldview as operational policy.

Map the warehouse boundary and reclassify accordingly. Document exactly where your governance infrastructure ends and where agents begin to operate. That gap is your liability exposure. Closing it requires semantic infrastructure: ontologies defining meaning, knowledge graphs encoding rules, provenance tracing decisions. It also requires reclassifying the investment. If your AI program reports to procurement or IT operations, it is structurally misclassified. Agentic AI is an organizational transformation requiring tolerance for failure, learning loops, and capability measurement. Budget accordingly.

Assign a single accountable owner for the illusion gap. If no one in your organization is responsible for the distance between what demos show and what infrastructure supports, the gap will widen with every pilot. This is CDO territory, but only if the CDO has authority over semantic infrastructure, not just data pipelines. If the CDO does not bridge data management and knowledge management, no one can write the contract that makes agents governable. In every organization I have assessed, the only function structurally positioned to own this gap is the CDO, if the role is empowered, and if it survives. Most do not. Average tenure has dropped below three years, often because the illusion gap makes the role appear redundant precisely when it is most needed.

The question for your board is not whether the demos look impressive. They will. The question is whether the infrastructure beneath them can survive production. The organizations that close the illusion gap now will capture disproportionate value when they scale. The ones that do not will discover the gap the hard way: in production, under load, with the board’s name on the liability. Approving interfaces without foundations is not a technology decision. It is a breach of governance.

Next In This Series

Part 2d examines the organizational crisis hiding behind the illusion layers. The CDO role, the only function positioned to bridge data management and knowledge management, is being eliminated at the moment it is needed most. Average CDO tenure has dropped below three years. The role most prepared to close the governance gap is the role most at risk.

The article also examines where AI actually works, where it breaks, and why. The evidence from enterprises that have succeeded, and the pattern of where LLM-generated shortcuts fail, reveals what boards must understand about the jagged frontier between success and failure. The board framework that closes Arc 2 gives directors the diagnostic tools to distinguish organizations building durable advantage from those accumulating hidden liability.

It publishes in two weeks. Subscribe to follow the series.

Subscribe now

About The Author

Frédéric Verhelst helps leadership teams build the foundations for trustworthy agentic AI. With a PhD in Applied Physics and twenty-five years at the intersection of data, AI, and industrial operations, he has led large semantic interoperability programs across energy companies, driven digital twin adoption at enterprise scale, and advised on governance architecture where autonomous systems meet strategic risk management. He works at Viking Life-Saving Equipment, where agentic AI governance for mission-critical safety operations is taking shape.

Follow him on LinkedIn for the latest posts in The Ontology Imperative: Building Trustworthy Agentic AI.

2b – The Stakes and the False Solutions – Why AI Investment Without Foundations Fails and Why LLMs Cannot Close the Gap

Frédéric Verhelst — Thu, 26 Feb 2026 08:15:09 GMT

Reading time: ~26 minutes

Summary

Your board approved AI investment. Most of it will fail. Not because the models are weak, but because the foundations are missing. Three paradoxes define the crisis.

Adoption rises while earnings stay flat.
Trust collapses while authority expands.
Cancellation rates increase while budgets grow.

European regulators warn of correlated failures that scale faster than human intervention can respond. The 737 MAX parallel is not a metaphor. It is structural.

The two solutions dominating boardroom conversations, observability dashboards and better prompts, cannot close the gap because both are detective controls. They operate after the fact. Agentic AI requires preventive architecture. When the tactical fixes are exhausted, leadership reaches for a final shortcut and asks whether LLMs themselves can stand in for governance. The evidence is clear. They cannot. LLMs are pattern engines, not reasoning systems. They cannot define authorization or trace decisions to authoritative sources.

But the problem runs deeper than reasoning failure. Training is compression. Compression strips attribution. Parametric outputs lack chain of custody. LLMs do not merely fail to provide provenance. They systematically destroy it. Every document ingested during training loses its attribution graph. Every citation risks fabrication. Every knowledge connection between a claim and its source is severed at industrial scale. Information scientists call this knowledge network decay. It is not a bug. It is the architecture. Without explicit provenance infrastructure, board oversight becomes unverifiable.

Organizations that understand this distinction will build the semantic infrastructure that makes agentic AI governable. The rest will explain failures to regulators. The potential is extraordinary. The foundations are non-negotiable.

Board Brief: Three Numbers and a Structural Warning

78% of companies use AI in at least one function. 80% report no material impact on earnings. Over 40% of agentic AI initiatives will be cancelled by 2027. These numbers come from McKinsey, Harvard Business Review and Gartner. Together, they describe adoption without value, delegation without trust and investment without governance.

Regulators see the same pattern. In December 2025 the European Systemic Risk Board identified eleven systemic risks that intensify when AI shifts from advising to acting. Model uniformity produces correlated failures. Speed eliminates the human intervention window. Opacity prevents explanation.

Two vendor-promoted solutions dominate boardroom conversations: observability tools and prompt engineering. Neither provides governance. Observability detects failures after execution. Prompts fail under adversarial pressure by mathematical proof. A third shortcut, asking LLMs to generate their own governance structures, fails quantitatively: peer-reviewed research shows LLM-generated ontologies achieve property alignment scores below 0.27 on a scale of 1.0.

A fourth dimension compounds all three: LLMs do not merely fail to produce governance. They actively destroy the provenance infrastructure governance requires. Peer-reviewed research documents that data authenticity, consent, and provenance for AI are fundamentally broken. GPTZero’s analysis of 4,841 NeurIPS 2025 accepted papers found over 100 fabricated citations across 51 papers that passed peer review. Even at roughly 1%, contamination enters the scholarly record and recycles into future training data. When your agents make decisions based on knowledge stripped of its chain of custody, every output is an assertion without evidence.

The Systemic Risk Audit: four questions for your next board meeting.

The Deployment Check: What percentage of our AI investment is in production rather than pilots? Action: Request a Production Readiness audit for all active pilots.
The False Solutions Check: Can we distinguish detective controls from preventive controls in our architecture? Action: Direct the CTO to map preventive controls versus monitoring across all agent deployments.
The 737 MAX / Regulatory Check: If a regulator asked for the authorization chain behind an agent decision, could we provide it within 24 hours? Action: Require a live demonstration of a decision trace for the top three high-risk agents.
The Provenance Check: Can we trace the provenance of the knowledge our agents use to make decisions, or are they operating on assertions without evidence? Action: Review the chain of custody policy for all third-party and internal training data.

If any answer is uncertain, the gap in this article is yours.

Executive Brief

Part 2a established the contract gap: most boards cannot answer three basic questions about how their agents make decisions. The contractor framework showed why prompts are not contracts and why GenAI governance cannot scale to agentic AI. The evidence demonstrated that prompt-based safety fails by design, not by accident.

Part 2b reveals the cost of that vacuum: a triple paradox of failing investment, systemic risk that regulators are already flagging, and a provenance crisis that is quietly dismantling the corporate record.

This article escalates the stakes. The gap is not academic. It is measured in billions of unreturned investment, regulatory warnings from the highest levels of European financial oversight, and cancellation rates that will consume nearly half of agentic AI initiatives within two years. When organizations reach for solutions, they find two that sound reasonable and one that sounds clever. All three are insufficient.

I have watched this pattern before. In the energy sector, organizations deployed autonomous operations systems with monitoring infrastructure and operational procedures but without semantic foundations. The monitoring worked perfectly. It recorded every failure with precision. What it could not do was prevent the failure from occurring. The organizations that built ontologies first, encoding decision boundaries before granting agents authority, operated at scales the monitoring-first organizations never reached. The architecture determined the outcome. The technology was identical.

The same dynamic is playing out across every industry now deploying agentic AI. The technology is powerful. The governance architecture is absent. And the solutions being sold do not address the structural problem.

I believe in agentic AI. I have built multi-agent systems grounded in formal ontologies since 2006. When I led the Integrated Operations in the High North program coordinating 23 energy companies, the organizations that encoded decision boundaries in ontologies before granting agents operational authority were the ones that reached production scale. The ones that deployed first and planned to add governance later accumulated technical debt faster than their teams could address it. The technology was identical across both groups. The architecture was not.

Today I am applying these same governance principles to multi-agent systems in mission-critical safety operations, where the cost of an ungoverned agent decision is not a financial loss but a human one.

The potential is real. But potential without foundations produces the paradoxes this article documents: faster but more fragile, more autonomous but less accountable, more scalable but more brittle.

The question is not whether to deploy agentic AI. The question is whether to build the infrastructure that makes deployment safe, or to discover the gap when something breaks publicly.

The Triple Paradox: Why Billions Produce So Little

Three paradoxes define enterprise AI in 2025-2026. Each is well documented. Together, they reveal a single root cause.

The Triple Paradox. Three data points from three independent research firms converge on a single diagnosis: enterprise AI is scaling investment without building foundations. Sources: McKinsey Global Survey 2025, Harvard Business Review Analytic Services 2025, Gartner 2025-2026.

The Deployment Paradox

McKinsey’s 2025 Global Survey delivered a finding that should have triggered emergency board meetings across every industry: eight in ten companies now use AI, most with generative AI capabilities. The same proportion report zero material impact on earnings.

This is not early-adoption lag. This is structural failure at scale.

Most deployments are horizontal: chatbots, copilots, email summarizers. These are efficiency tools that live in the 18% of value I helped quantify in the OLF/Epsis Integrated Operations Value Assessment. The study showed that 82% of digital transformation value comes from entirely new capabilities, transformation rather than optimization. Chatbots optimize. Agentic AI transforms. But transformation requires foundations that optimization never demanded.

Vertical use cases, the ones that create competitive advantage, remain stuck in pilot. Fewer than 10% reach production. Boards celebrate adoption metrics while ignoring production economics. That is not progress. That is theatre with a technology budget.

The Trust Paradox

Harvard Business Review surveyed 600 technology leaders in mid-2025. The finding defines the current crisis: investment rises while trust collapses.

Only 6% trust AI agents with core business processes. 43% restrict agents to routine tasks. 39% allow only supervised, non-core use cases. Yet MIT Sloan projects 250% growth in AI decision-making authority over the next three years.

Read those numbers together. Budgets expand. Authority scales. Trust evaporates. Organizations are simultaneously spending more on AI and trusting it less. That is not a paradox of perception. It is a signal that the governance infrastructure required to justify trust has not been built.

Capgemini’s C-suite research reinforces this from a different angle: 67% of executives say governance protocols are the only way to safely use AI for decisions. Only 34% have them. The gap between knowing what is needed and building it is where the real risk accumulates.

The Trust-Authority Divergence. Authority delegated to AI agents is scaling exponentially while executive trust collapses. The widening gap is the governance infrastructure that has not been built. Sources: MIT Sloan 2025, Harvard Business Review Analytic Services 2025.

The Cancellation Paradox

Gartner predicts that over 40% of agentic AI projects will be cancelled by 2027 due to cost, unclear value, and weak risk controls.

Gartner also identified a phenomenon they call agent washing. Vendors rebrand chatbots and robotic process automation as agentic AI. Of thousands of solutions claiming agentic capabilities, only about 130 are genuine. Agent washing is not a buzzword problem. It is systemic distortion. Boards are funding the illusion of autonomy at enterprise scale while the underlying capabilities remain unchanged.

The cancellation paradox is the deployment paradox reaching its conclusion. Pilots that never reached production become budget lines that never produced returns. The investment was real. The foundations were not.

One Root Cause

Three paradoxes. One explanation.

The deployment paradox exists because there is no semantic infrastructure connecting AI actions to measurable business outcomes. Without ontologies that define what success means in machine-readable terms, organizations cannot measure whether agents are producing value or merely producing activity. Architecture is not just for safety. It is the only way to move AI from a cost-centre experiment to an autonomous operations revenue driver.

The trust paradox exists because there are no enforceable governance controls justifying the authority being delegated. Trust is not a sentiment. It is a structural property of systems that can demonstrate accountability. The 6% who trust agents with core processes are not more optimistic than the 94%. They have built more infrastructure.

The cancellation paradox exists because organizations are scaling experiments without the foundations production requires. Pilots succeed in narrow scope because the scope itself provides implicit governance. The team running the pilot understands the context, reviews the outputs, and intervenes when something goes wrong. Production removes that implicit governance. Scale removes the human safety net.

The World Economic Forum’s 2025 report on AI agent governance warns of orchestration drift, semantic misalignment, and cascading failures across multi-agent systems. These are not theoretical edge cases. They are enterprise risks in motion.

This is the same pattern I saw as CEO of the Knowledge Graph Alliance. Airbus, Michelin and Bosch worked in different industries, but the failure modes were identical. Monitoring existed. Procedures existed. Semantic foundations did not. The technology was identical. The outcomes were not.

CFO advisor Matteo Turi frames it with the precision boards need: the wrong question is which AI tool should I use. The right question is how should I organize my business processes so that AI can scale them. ChatGPT and AI tools do not move the needle. AI agents do. But agents need architecture, not tools.

AI does not create wealth. Architecture does.

The 737 MAX Parallel: When Governance Gaps Become Systemic Risk

The 737 MAX crisis was not a technology failure. It was a governance failure. Boeing’s MCAS system worked as designed. The design itself was ungovernable: insufficient pilot training, inadequate override procedures, regulatory capture that allowed self-certification. The technology functioned. The accountability architecture was absent.

Agentic AI is heading down the same path. Not because the models are dangerous, but because the governance structures required for autonomous systems have not been built.

In December 2025, the European Systemic Risk Board published a report identifying eleven systemic risks from AI in financial markets. Each risk amplifies when AI shifts from advisory to autonomous operation.

Model uniformity is the risk that keeps regulators awake. When organizations deploy similar AI models, failures become correlated. They cluster in time. They cascade across systems. This is the 2008 financial crisis lesson applied to AI: risk models that assume independence fail catastrophically when exposures move together. A dozen banks using the same foundation model to make lending decisions will make the same mistake simultaneously. The portfolio of AI decisions is not diversified. It is concentrated.

Excessive trust compounds the uniformity problem. Early success breeds overconfidence. Boards delegate more authority to agents. Oversight declines precisely when risk is scaling.

Speed eliminates the human intervention window. Autonomous systems act faster than humans can review. By the time an alert reaches a decision-maker, the agent has already executed with corporate authority. The damage precedes the notification.

The ESRB report identifies a risk dynamic that most technology discussions miss entirely: procyclicality. When AI agents optimize for the same signals using similar models, they amplify market movements rather than dampening them. Individual optimization produces collective fragility. Each agent acts rationally within its own scope. The aggregate behavior is irrational at system level. This is not a failure of any single agent. It is an emergent property of architecturally ungoverned systems operating in the same environment.

The governance response to procyclicality requires something no individual model can provide: awareness of system-level constraints that transcend the agent’s operational scope. An ontology can encode these constraints. A knowledge graph can represent the relationships between agents, markets, and boundaries. Guardrails can enforce limits that no individual agent would choose to impose on itself. This is precisely the governance architecture that observability tools and prompt engineering cannot deliver, because the problem is not what any single agent does, but what all agents do together when no structural constraint coordinates their behavior.

Beyond Financial Markets

The ESRB focuses on finance, but the dynamics threaten any sector with autonomous systems operating in real-time environments. Energy grids. Commodity trading. Supply chains. Healthcare. Telecommunications.

Researchers simulated AI-enabled power grid operators. Performance improved until a critical mass adopted AI. Then the grid destabilized. Operators without AI could not interpret the unexpected moves of those with it. The individual optimization produced collective fragility.

The Bank of England notes that only 2% of AI use cases are currently fully autonomous. That percentage is about to scale dramatically. The window to build governance foundations is open now. It will not remain open indefinitely.

Capital buffers, enhanced supervision, and liquidity requirements are necessary regulatory responses. But they are insufficient. Regulators can require preparation for failure. They cannot provide the architecture that prevents it. That architecture, ontologies, knowledge graphs, guardrails, provenance, is the organization’s responsibility.

Capital buffers are seatbelts on a plane without wings. Architecture is the wing.

False Solution One: Observability Tools

The observability market exploded in 2025. Fiddler AI, DataRobot, Superwise, Arize AI. Every platform promises transparency and trust. The positioning is compelling: see what your agents are doing, catch problems early, maintain control.

The promise is real but incomplete. And the incompleteness is where the danger lives.

IBM research establishes the boundary clearly: observability tools flag anomalies after they occur. They cannot explain why. The decision path remains opaque. Observability shows what happened and when. Not why this decision. Not who authorized it. Those answers require governance architecture no dashboard can deliver.

Microsoft Azure’s own framework for agent observability specifies five requirements: metrics, traces, logs, evaluations, and governance. Most vendor platforms deliver four of the five. Governance is the missing component because governance is not a feature. It is architecture.

Detective Controls Are Not Preventive Controls

The distinction between detective and preventive controls is fundamental in financial governance. Boards understand it intuitively in every domain except AI.

A monthly audit that catches a $50,000 unauthorized expense is a detective control. An approval workflow that blocks the unauthorized transaction before it executes is a preventive control. No CFO would accept monthly audits as a substitute for authorization workflows. Yet boards routinely accept observability dashboards as substitutes for AI governance architecture.

Detective Controls vs Preventive Controls. Observability detects failures after execution. Governance architecture prevents them before the agent acts. The distinction is structural, not a matter of configuration. No dashboard upgrade converts a detective control into a preventive one.

GenAI hallucinates and you fix text. Agentic AI hallucinates and acts. When an observability alert arrives on your phone, the agent has already executed with your authority. You are not preventing failure. You are documenting it.

I saw this boundary at TotalEnergies EP Danmark. As Head of Data Office I oversaw cross-domain prioritization and ensured business alignment as teams developed the semantic virtualization connecting drilling, reservoir engineering and production data. A single concept, “well,” existed in the production database, in drilling reports, in proprietary reservoir models, and in regulatory filings. Each system defined it differently. The semantic layer gave it one truth. Before that layer matured, monitoring captured every signal with precision. It could tell you exactly what had failed, and where, and when. What it could not do was reason across domains to detect the conditions producing that failure before it arrived. After the semantic foundations were established, definitions and relationships became explicit. Systems could connect a pressure anomaly in a drilling report to a reservoir model constraint to a regulatory threshold, not because the monitoring improved, but because the architecture allowed meaning to travel across boundaries. The data did not change. The architecture did.

Observability is forensics. Governance is authority itself. One tells you what went wrong. The other prevents it from happening.

Sophisticated Monitoring Is Still Not a Safety Net

Microsoft’s Model Context Protocol claims observability at every layer, embedding governance in the connectivity standard. The fine print matters. MCP is a connectivity standard between agents and data sources. It enforces only what you build on top of it. Without ontologies defining business meaning, MCP delivers information, not authority. Connection is not governance.

Specialized AI safety monitors represent the next generation of observability. They detect risks after agents propose actions but before execution. This sounds preventive. It is not. These monitors watch. They can flag. They cannot enforce structural constraints. They operate within the probability space of the model they are monitoring. A monitor built on the same architectural foundation as the agent it supervises inherits the same limitations.

We do not need better monitors. We need contracts machines cannot talk their way around.

Boards buying observability without ontologies, knowledge graphs, and authorization controls are funding theatre. You will detect failures with precision. You cannot prevent them.

False Solution Two: Better Prompts

Prompt engineering defined 2023 and 2024. For generative AI, it worked. You crafted better instructions, got better outputs, reviewed them, and published. The human review step made prompt imprecision tolerable.

Then agentic AI arrived, and the math closed in.

UC Berkeley, Stanford, and LMU Munich researchers proved that prompt-based safety cannot work at scale. The formal argument is straightforward: a weaker filter cannot reliably separate harmful prompts from harmless ones when the model it is filtering is more capable than the filter itself. This is not an engineering limitation that better prompting techniques will overcome. It is a theoretical limit.

Cassie Kozyrkov, Google’s former Chief Decision Scientist, says it plainly: you cannot secure a strong model with a weaker filter.

OWASP, the foundation that sets security standards for the software industry, ranks Agent Goal Hijacking as the number one AI security risk. The mechanism is indirect prompt injection: instructions embedded in documents, web pages, or retrieval-augmented generation context that agents ingest and treat as actionable guidance. These injections do not need to be hostile. Ordinary content can be misinterpreted as commands. The agent’s objectives get redirected, and the redirection cascades through its toolchain, executing with corporate authority.

Recent jailbreak research confirms this is not an engineering problem awaiting a fix. It is a structural property of how language models work. Berkeley’s IRIS attack demonstrated universal and transferable adversarial suffixes that bypass safety alignment across frontier models, including those from OpenAI, Google, Meta, and Anthropic. HiddenLayer’s Policy Puppetry technique achieved the same result through a different mechanism: a single prompt template that exploits how models process policy-related instructions, working across all major frontier providers. These are not isolated exploits. They are demonstrations that prompt-level safety cannot enforce constraints once agents can act. The attacks transfer across model architectures, across providers, and across model generations. Newer models are not inherently safer.

Production data confirms this is no longer theoretical. In a single week in early 2026, one monitoring service detected over 28,000 threats across 74,000 agent interactions. Indirect and instruction-style injections accounted for more than 6,000 events. RAG poisoning reached 10%. Inter-agent attacks emerged as an entirely new category. Goal hijacking, the scenario OWASP ranks as the top risk, was measured in real production traffic with 97% detection confidence. These are not projections. They are measurements from systems operating under corporate authority right now.

The Prompt Safety Structural Limit. When the model’s capabilities exceed the filter’s ability to constrain them, safety becomes probabilistic rather than enforceable. Universal jailbreak attacks now transfer across architectures, providers, and model generations.

Unlike database vulnerabilities where security teams could sanitize inputs, prompt injection exploits meaning itself. Your security perimeter just expanded from user inputs to every document, email, database record, and web page the agent reads. Traditional firewalls do not catch meaning. Semantic infrastructure does, because it defines what meaning is valid in your operational context and rejects everything else.

Prompts fail by design, not by accident. Observability reports failures you could have prevented. Neither is governance.

False Solution Three: Let the LLM Write the Ontology

When organizations recognize they need ontological foundations but lack the talent to build them, a seductive shortcut presents itself: ask the LLM to generate the ontology. The logic seems sound. LLMs excel at language. Ontologies use language-like constructs. Why not let it produce one?

Peer-reviewed research provides the answer quantitatively.

Plu et al. presented results at the International Semantic Web Conference in 2024 evaluating LLM-generated ontologies against expert-created benchmarks. The findings are definitive: LLMs achieve roughly 70% alignment on class identification, plausible naming of concepts. But property modeling, the structural relationships that give ontologies their reasoning power, scores a maximum F1 of 0.27 on a scale of 1.0. The researchers characterized the output as shallow, flat design. The LLM names things plausibly. It cannot model their relationships formally.

Fathallah et al. confirmed these findings at the European Knowledge Acquisition Workshop the same year, documenting that LLM-generated ontologies consistently lack hierarchical depth. The models produce flat structures that look like taxonomies rather than the rich axiomatized models that formal reasoning requires.

This is not a training data problem that will resolve with larger models. It is a capability boundary. Ontology engineering requires the abstraction skills of formal modeling: deciding what to include and what to deliberately exclude, capturing phenomena of interest while neglecting lower-order effects, encoding constraints that survive edge cases the modeler has never seen. That judgment is not pattern matching. It is domain expertise encoded as formal logic.

Skilled ontologists increasingly use LLMs to work faster, and this is not a contradiction. LLMs can surface candidate terms, extract vocabulary from thousands of documents and generate early sketches worth evaluating. They are accelerators, not architects. The engineering judgment that decides what belongs in a domain model, what must be excluded and how constraints survive edge cases remains human. The model can propose. The human decides. Governance stays with the people who understand both the business and the consequences of getting the logic wrong.

You will hire an ontologist either at the beginning or six months later when the LLM-generated ontology fails under production pressure. The timing is optional. The need is not.

Why LLMs Are Not Reasoning Engines: The Board-Level Argument

Your board just saw a demo of a reasoning model that thinks before it acts. The narrative was seductive: clean logic, structured chains, step-by-step reasoning. The reality is that these models look like they reason. They do not.

This is not a contrarian position. It is the consensus of the researchers who build these systems.

What LLMs Actually Do

Andrej Karpathy’s 2025 review documented what he calls jagged intelligence. The same model that solves advanced mathematics fails at basic logic. The capability profile is not a smooth gradient from easy to hard. It is unpredictably jagged, brilliant in one domain, incompetent in the adjacent one.

Yann LeCun, formerly Meta’s Chief AI Scientist, explains the mechanism. LLMs do not build abstractions. They do not test hypotheses. They do not maintain persistent world models. They predict the next token based on statistical patterns in training data. Their apparent reasoning is pattern continuation dressed as thought.

Reinforcement Learning from Verifiable Rewards appeared to address this limitation. Models trained with RLVR produce outputs that look more like structured reasoning. Karpathy clarified the distinction: RLVR rewards behaviors that correlate with correct answers. It does not produce understanding. It produces optimization. The model learns to produce reasoning-shaped outputs because those outputs receive higher rewards, not because the model has developed the capacity to reason.

Pattern optimization is not reasoning. And it will never be governance.

Virginia Dignum, Professor of Responsible AI at Umeå University and co-chair of the ACM Technology Policy Council, frames the institutional consequence. AI systems are designed artefacts, not autonomous actors. They lack intentions, goals, and agency. The risk is not machine consciousness but institutional over-reliance on systems that cannot understand what they are doing. That is a governance challenge, not an existential one. In The AI Paradox she argues that the future depends on cooperation and institutional design, not superintelligence. Framing AI as unstoppable destiny weakens agency and diffuses accountability. It is abdication dressed as analysis.

The debate about whether AI will someday reason is a distraction from the fact that it does not reason now, and boards are deploying it as if it does.

The Capgemini Test

Capgemini demonstrated the practical consequence in their Data-Powered Innovation Review. They asked an LLM which subsidiaries of a fictional company had ESG contributions. The source document listed two subsidiaries with ESG contributions and one without.

The LLM’s answer was fluent and wrong. It confidently attributed ESG contributions to SubZ, a subsidiary that had none in the source material. The model invented a plausible answer because plausibility is what language models produce. Without grounding, plausibility beats truth every time.

Then Capgemini converted the same document into a knowledge graph: nodes, edges, structured relationships. When the LLM queried the graph using a formal query language, it returned only the subsidiaries with actual ESG contributions. No hallucinated SubZ.

The model did not improve. The graph constrained reality. The LLM was the same LLM. The architecture was different. That is the entire argument in one experiment.

The same LLM querying the same information produces a hallucinated answer from documents and an accurate answer from a knowledge graph. The model did not improve. The graph constrained reality.

When you query a document index, you validate every answer. When you query a knowledge graph, you validate the graph once. That distinction determines whether your AI architecture scales with human oversight, which it cannot at production volumes, or scales with the quality of your semantic infrastructure, which it can, indefinitely.

What Boards Need to Understand

When an agent moves $2M at 14:23 UTC, regulators will not ask how the model explained the action. They will ask what authorization permitted it.

LLMs cannot answer this question. They lack persistent memory. They lack structured representation. They lack rule enforcement. They lack provable authorization chains. They explain actions fluently. They do not govern them.

The fiduciary consequence is arriving faster than the governance infrastructure. AI-related securities class actions doubled from 2023 to 2024, with average D&O settlements reaching $56 million. The legal standard is clear: plaintiffs do not need to prove the AI system failed. They need to prove the board failed to govern it. Under the Caremark doctrine, failure to implement a functioning oversight system is itself a breach. The EU AI Act introduces two novel fiduciary duties: AI due care and AI loyalty oversight. Across jurisdictions, courts converge on a single principle: omission is liability. Yet the NACD's 2025 survey found only 6% of boards have established AI-related management reporting metrics. The gap between fiduciary obligation and governance reality is where personal liability accumulates.

Observability shows what happened. Reasoning traces show why the model thinks it happened. Governance constrains what may happen. These are three different capabilities. Only the third prevents unauthorized actions.

Your board needs three layers no LLM can supply:

Ontologies that define what authorized means in your specific business context.

Knowledge graphs that encode the boundaries of what agents may decide.

Provenance systems that trace every decision to a specific authority before execution.

These are preventive controls. They operate before risk manifests. They do not depend on the model’s capability to reason. They constrain the model’s ability to act regardless of what its reasoning produces.

The Provenance Crisis: Why LLMs Destroy the Knowledge They Depend On

The previous sections establish that LLMs cannot reason and cannot build governance structures. But the problem is worse than capability failure. LLMs do not merely fail to provide provenance. They architecturally destroy it.

Jessica Talisman, an information architect with 25 years of experience in semantic engineering, has documented this destruction with a precision that boards need to understand. She traces provenance as a foundational principle of reliable knowledge, one that has been maintained for nearly two centuries through archival science, library systems, and institutional memory. LLMs are the first technology to systematically dismantle it at industrial scale. Her term for this phenomenon, knowledge network decay, describes a dynamic consistent with the peer-reviewed literature on broken AI data provenance but names what boards need to recognize: not a technical flaw, but a structural dissolution of the connections between claims and their sources.

How LLMs Destroy Provenance

To understand the destruction, distinguish between two layers inside every large language model. The first is the retrieval layer: search results, RAG pipelines, uploaded documents. When an LLM cites a source, it draws from this layer. But retrieval only records where the model looked at query time. It says nothing about how the model arrived at its interpretation.

The second layer is where the damage occurs. The parametric layer, the model’s trained weights, is built through compression. Billions of documents are reduced to statistical distributions across billions of parameters. What survives is pattern. What does not survive is provenance. The chain connecting a claim to its author, a definition to its standard, a finding to its methodology, all of it is stripped during training. As Talisman documents, the entire relational infrastructure that knowledge systems exist to maintain is dissolved in the process. Metadata does not travel with the weights. Attribution does not survive the pipeline.

This matters because the compression is irreversible. It is not a limitation that scaling or fine-tuning will resolve. The architecture requires it. A model that preserved full provenance for every ingested document would not be a language model. It would be a database. The parametric layer is, as Talisman puts it, a provenance-free zone, not by oversight, but by design.

Documents enter training with metadata intact: author, date, source, version. Compression reduces them to statistical distributions. What emerges has no chain of custody. The retrieval layer can show where the model looked. It cannot show where its understanding originated. The parametric layer is, as Talisman describes it, a provenance-free zone.

I recognized this pattern immediately because I had seen it before in a different domain. In reservoir engineering, before the adoption of Ensemble Kalman Filter methods and Fast Model Update techniques, organizations relied on single “best estimate” models. These models collapsed an entire uncertainty envelope into one number. They were precise, confident, and routinely wrong, because the compression destroyed the very information that would have signaled when the estimate was unreliable. When organizations at Shell, Equinor, and TotalEnergies adopted ensemble methods, they stopped compressing uncertainty into a point estimate and started maintaining the full distribution. The models did not improve. The architecture preserved information that compression had destroyed.

LLMs do to knowledge what best-estimate models did to uncertainty. They compress provenance into statistical patterns, producing outputs that are confident, fluent, and untraceable. The information that would allow you to evaluate whether an output is reliable, where the claim originated, what evidence supports it, what context qualified it, does not survive training. Your business lives in the edges. Compression amplifies the mean.

Vibe Citing: When Fabrication Mimics Legitimacy

The provenance destruction has a particularly dangerous downstream effect. When LLMs generate citations, they do not retrieve references from a database. They predict what a citation should look like based on statistical patterns. The result is what GPTZero’s research team calls vibe citing: fabricated citations with structural markers of legitimacy that look correct at first glance but collapse under scrutiny.

GPTZero confirmed over 100 fabricated citations across 51 accepted NeurIPS 2025 papers out of 4,841 scanned. These papers had each been evaluated by at least three peer reviewers. The fabrications ranged from obvious placeholders to sophisticated amalgamations of real author names, journal titles, and DOI formats that pointed to nothing. Every one of them passed review under volume pressure. Even at roughly 1%, contamination enters the scholarly record and recycles as training data.

This is not an academic curiosity. It is a governance catastrophe in formation.

When vibe-cited papers enter the scientific record, they become training data for the next generation of models. The fabricated citations acquire the structural markers of legitimacy: they appear in peer-reviewed venues, they get indexed by search engines, they get cited by other papers. Each cycle further dilutes the connection between claims and their original sources. Researchers call the statistical dimension of this problem model collapse. The knowledge dimension is worse. It is provenance decay compounding across generations.

LLM-fabricated citations pass peer review, enter the scholarly record, get indexed and cited, and become training data for the next generation of models. Each cycle compounds provenance decay. The mechanism is recursive and self-reinforcing.

Now translate this to your enterprise. Your AI agents make decisions based on knowledge retrieved from documents, databases, and web sources. If that knowledge was generated or processed by LLMs at any point in the chain, the provenance may already be compromised. The agent is not operating on evidence. It is operating on assertions that mimic evidence. When a regulator asks where a decision came from, the chain of custody may be fabricated from the start.

From Agent Risk to Ecosystem Risk

Previous sections argue that individual agents need governance: ontologies defining scope, knowledge graphs encoding boundaries, guardrails constraining behavior. All of that remains true. But the provenance crisis adds a dimension most governance frameworks miss entirely.

Your agents do not operate on raw data. They operate on knowledge, on claims, relationships, and inferences derived from documents, databases, and retrieval systems. If the knowledge infrastructure feeding your agents has been contaminated by provenance-free content, if fabricated citations have entered your retrieval pipelines, if LLM-generated summaries have replaced sourced analysis in your corporate knowledge base, then your governance architecture is built on sand regardless of how well designed it is.

The MIT Data Provenance Initiative led by Longpre et al. documented the scale of this problem: over 70% of AI training datasets have unspecified licenses, and the provenance chains connecting datasets to their original sources are routinely broken or absent. Data authenticity, consent, and provenance for AI are fundamentally broken. This is not a future concern. It is a current condition.

Governance requires provenance at two levels. First, decision provenance: tracing every agent action to the specific rules, data, and authorization that permitted it. This is the focus of the Decision Contract Checklist from Part 2a. Second, knowledge provenance: verifying that the information feeding agent decisions is itself traceable to authoritative sources. Without both levels, your governance architecture has a foundation problem.

The Decision Contract Checklist’s Evidence Pack, execution-time provenance recording which rules were evaluated, which sources were consulted, which exceptions were considered, is necessary but not sufficient. If the sources themselves lack provenance, the Evidence Pack documents a process built on unverified assertions.

This is where W3C PROV, the provenance ontology standard, becomes operationally critical. Not as documentation. As infrastructure. Design-time: facts in the knowledge graph carry PROV links to their origins, connecting every assertion to the data source, transformation process, and responsible agent that produced it. Run-time: each agent action records which graph facts, rules, and sources were evaluated during execution. Together, these two layers allow organizations to verify not just that an agent followed its rules, but that the knowledge informing those rules traces to authoritative origins.

The irony is architectural. Tim Berners-Lee’s Semantic Web, proposed in 2001, was designed precisely as provenance infrastructure for the digital age. RDF, ontologies, URIs, the entire W3C standards architecture was built to maintain the chain of custody between claims and their sources. Using a different URI for each specific concept prevents semantic collapse. The Semantic Web was trust architecture. LLMs are the structural opposite.

I was building this provenance infrastructure in 2008, when I presented at the W3C Semantic Web in Energy Industries workshop in Houston, demonstrating how ontologies could maintain knowledge integrity across complex industrial operations. The AI industry was still treating knowledge as disposable training data. Seventeen years later, the infrastructure we built then is exactly what agentic AI governance requires now.

Provenance is how institutions remember. Without it, neither agents nor humans can be accountable.

The Architecture Your Board Assumes Exists

The pattern across all three false solutions and the provenance crisis is identical. Each operates at the execution layer, not the accountability layer.

Observability tracks actions. It does not define what agents must never do.

Prompts guide behavior. They cannot enforce constraints under adversarial pressure.

LLM-generated structures name concepts plausibly. They cannot model the formal relationships that governance requires.

LLM-processed knowledge strips provenance. It cannot maintain the chain of custody that accountability demands.

Agentic AI is not generative AI. GenAI hallucinates and you fix text. Agentic AI hallucinates and acts. When the false solutions fail, and the evidence shows they will, the agent has already executed with your authority.

Real governance requires structural infrastructure:

Formal ontologies encoding business rules as machine-readable logic. Not documentation that describes rules, logic that enforces them. When an ontology defines approved supplier, it specifies exactly what qualifications are required, what exceptions exist, what conditions must be met. The agent cannot reinterpret the definition.

Knowledge graphs where terms have precise, consistent meanings across the organization. When one division’s approved supplier differs from another’s, the graph maintains both definitions and the rules governing each. Ambiguity that humans resolve through context becomes explicit and computable.

Guardrails that agents cannot ignore or bypass. Hard constraints encoded in the system architecture. When an agent approaches a boundary, it stops or escalates. It does not rationalize its way past the limit. The distinction between soft guidance and hard constraints is the distinction between aspiration and governance.

Provenance systems tracing every decision to its authoritative source, and verifying that the knowledge informing those decisions is itself traceable to authoritative origins. Design-time provenance links facts to their sources. Run-time provenance records which facts and rules were consulted during execution. When something fails, you trace the chain. When a regulator asks, you show the chain. When a board member demands accountability, the chain provides it. When a fabricated citation has entered your knowledge base, the provenance system detects it before an agent acts on it.

Four layers of preventive infrastructure. No single layer is sufficient. Together, they constitute the defense-in-depth that boards assume exists.

This infrastructure shifts the human role from reviewer to designer. Human-in-the-loop means checking every output, a model that scales linearly with headcount and breaks at production volumes. Human-above-the-loop means designing the constraints, boundaries, and escalation criteria before the agent acts. The agent operates within those constraints autonomously. Review scales with headcount. Authority scales with design.

This is not theoretical. The W3C standards for provenance, formal semantics, and constraint languages are mature, with decades of implementation history. PROV-O provides the provenance ontology. OWL provides the formal logic. SHACL provides the constraint language. What is missing is organizational commitment to building infrastructure that does not show up on demo day.

When regulators come knocking, observability shows your agent moved $2M at 14:23 UTC. Governance proves the agent acted under Rule 7.3.2, triggered by an approved purchase order, within a defined spending limit, with a provenance trail connecting the action to a specific authorization, and with knowledge provenance verifying that the data informing the decision traces to authoritative sources rather than fabricated assertions.

One is forensics. The other is authority.

If your organization cannot show preventive controls, you do not have governance. You have a crime scene recorder.

What This Means for Your Board

The Triple Paradox, the systemic risk warnings, the three false solutions, and the provenance crisis converge on a single conclusion: the governance gap cannot be closed with the tools most organizations are buying.

Observability helps you operate agents. It does not make them trustworthy.

Prompt engineering helps you instruct agents. It does not make them safe.

LLM-generated ontologies help you name concepts. They do not make them governable.

LLM-processed knowledge helps you access information faster. It does not make that information reliable.

Governance Enables Speed

The argument for semantic infrastructure is not an argument against speed. It is an argument for sustainable speed.

The Capgemini ESG experiment is worth revisiting. The same LLM that hallucinated SubZ when querying documents returned perfectly accurate results when querying a knowledge graph. The graph did not slow the system down. It made the system trustworthy. The validation burden shifted from every answer to the graph itself, validated once, queried thousands of times. That shift is the difference between a system that scales linearly with human oversight and one that scales with the quality of its constraints.

The provenance argument reinforces this. When your knowledge graph is built on ontologies with formal provenance, every fact in the graph can trace its origin. The LLM querying the graph inherits the graph’s provenance without needing to provide its own. The architecture does the work that the model cannot. Validation happens once, at graph construction. Trust propagates through every query.

Ethan Mollick at Wharton argues that boards should treat AI foundations as R&D investment, not cost-cutting. The framing matters because it changes the time horizon and the success metrics. R&D investments are evaluated on capability creation, not immediate efficiency gains. The organizations treating semantic infrastructure as overhead will cancel it when quarterly pressure arrives. The organizations treating it as R&D will build the durable advantage.

Bret Taylor, co-founder of Sierra and former Salesforce co-CEO, advocates for defense-in-depth in AI architectures: multiple independent layers of protection rather than any single safeguard. Ontologies, knowledge graphs, guardrails, and provenance are precisely this: independent layers that each address a different dimension of governance. No single layer is sufficient. Together, they create the architectural depth that makes autonomous systems governable.

The infrastructure that closes the gap is not optional enrichment layered on top of AI deployment. It is the contract between your organization and the autonomous systems acting on its behalf.

The Investment Decision

The choice is structural: build foundations or fund failure.

Your next board meeting should not ask which AI model to deploy. It should ask what governance infrastructure has been built. The answer determines whether your AI investment produces transformation or joins the 80% that produced nothing.

OLF/Epsis IO Value Assessment, Norwegian Continental Shelf (2006). 82% of value came from capabilities that did not exist before. Most AI business cases still budget the 18%.

In a value assessment I contributed to on the Norwegian Continental Shelf, we quantified the split across 11 fields: 82% of value came from capabilities that did not exist before. Only 18% from cost reduction. Most AI business cases today repeat the same error. They budget for the 18%.

Reasoning models give you plausible answers. Governance architecture gives you trustworthy agents.

Which one is your board betting the company on?

Next In This Series

Part 2c examines board accountability and the organizational crisis that makes governance gaps persistent. The CDO survival crisis shows why the role most needed is the role most at risk. The dashboard trap explains why analytics maturity does not equal AI readiness. The board framework delivers five questions every director must ask about AI agents, the capstone that transitions from diagnosing problems to building solutions.

Arc 3 covers the architecture, governance, and implementation patterns that allow organizations to deploy agentic AI safely at scale. It introduces the Agentic AI Capability Stack and the CDO mandate that makes it operational.

It publishes in two weeks. Subscribe to follow the series.

About The Author

Frédéric Verhelst helps leadership teams build the foundations for trustworthy agentic AI. With a PhD in Applied Physics and 25 years at the intersection of data, AI, and industrial operations, he has led large semantic interoperability programs across energy companies, driven digital twin adoption at enterprise scale, and advised on governance architecture where autonomous systems meet strategic risk management. His perspectives on precision digital twins and semantic AI foundations were recently featured on the GraphRAG Curator Podcast. He works at Viking Life-Saving Equipment, where agentic AI governance for mission-critical safety operations is taking shape.

Follow him on LinkedIn for the latest posts in The Ontology Imperative: Building Trustworthy Agentic AI.

Acknowledgement

This article draws on engagement from practitioners, CDOs, and board advisors who commented on the LinkedIn series 'The Ontology Imperative.' Their challenges, corrections, and real-world examples sharpened the arguments presented here.

2a – The Missing Contract – Why Most Boards Cannot Govern What They Cannot Define

Frédéric Verhelst — Tue, 10 Feb 2026 09:43:12 GMT

Reading time: ~18 minutes

Summary

Most boards have approved agentic AI pilots. Few can answer three questions about how those agents make decisions: where the reasoning came from, whether it is genuine or hallucinated, and how to fix the root cause when something breaks. This article presents the evidence that the gap is structural, not operational, and that the solution requires accountability infrastructure most organizations have not built. The contractor framework, security research, and real-world governance models converge on one conclusion: you cannot instruct your way to trust. You must engineer it.

Thanks for reading The Ontology Imperative: Building Trustworthy Agentic AI!

Subscribe for free to receive new posts and support my work.

Subscribe now

Board Brief: Three Non-Delegable Decisions

AI agents act with corporate authority. When they fail, boards own the outcome legally, financially, reputationally. The Air Canada tribunal established this precedent: your agent is not a separate legal entity. It represents you.

Current governance approaches cannot prevent agent failures. Monitoring dashboards, safety prompts, and policy documents are reactive detective controls. Autonomous systems require proactive preventive architecture.

Governance requires semantic infrastructure most organizations have not built: ontologies defining agent scope, knowledge graphs encoding decision rules, guardrails preventing unauthorized actions, and provenance tracing every decision to its source. This gap is structural. It cannot be closed with vendor platforms or better prompts.

Three questions for your next board meeting:

Can we trace agent decisions to authoritative sources?
Can we distinguish reasoning from hallucination?
Can we fix root causes without retraining models?

If any answer is “no,” you have a governance gap that monitoring tools cannot close.

The Scale Unlock

Review scales with headcount. Authority scales with design.

The Board Audit

Most organizations cannot answer the three accountability questions. The gap is structural, not operational.

Executive Brief

Agentic AI is the most transformational enterprise capability I have seen in twenty-five years. Agents that can reason, plan, and execute autonomously will unlock value at scales we can barely imagine today: dynamic supply chain optimization responding to real-time signals, personalized customer resolution at enterprise scale, autonomous operational decision-making in complex environments. The organizations that deploy this capability effectively will create durable competitive advantages.

From 2006 to 2012 I worked on semantic programs, serving as Program Manager for the Integrated Operations in the High North (IOHN), a sidecar initiative to Equinor’s Global Operations Data Integration program, aligning operators and vendors on semantic web practices for autonomous operations. I returned to semantics in 2017, bringing ontology-first patterns into digital twin work and enterprise data initiatives, and later served as part-time and interim CEO of the Knowledge Graph Alliance, working with Airbus, Michelin, Bosch, TotalEnergies, universities, and standards bodies. Across all of these environments, the pattern was consistent: semantic contracts, not prompts, determine whether autonomous systems can operate with authority, scale safely, and deliver durable advantage. Where the contract layer was built up front, execution moved faster, scaled further, and carried less risk.

Boards do not approve prompts; they approve authority. The semantic contract is the authority.

This article focuses on LLM-based agentic AI and the hybrid architectures that combine foundation models with ontological grounding. These probabilistic systems face governance challenges that earlier agent architectures (symbolic, rule-based) did not: hallucination, non-determinism, and prompt injection vulnerabilities. The semantic infrastructure described here transforms soft constraints (prompts and filters) into hard constraints (domain axioms and real-time validation) that make these systems governable at scale.

The difference between winning and failing at agentic AI comes down to a single architectural choice: do you build the accountability infrastructure that allows agents to act with authority, or do you deploy them with prompts and hope?

When organizations deploy AI agents that approve transactions, reject suppliers, grant access, or send communications on corporate authority, three questions determine whether they can capture the transformation or just accumulate the risk:

Can you trace an agent decision to its authoritative source? Can you distinguish reasoning from hallucination? When it breaks, can you fix the root cause?

If the answer to any of these is no, we are delegating authority to probabilistic systems without accountability infrastructure. The gap is architectural. And the evidence shows it cannot be closed with better prompts, monitoring dashboards, or policy documents.

Parts 1a and 1b established why knowledge graphs provide competitive advantage and why most organizations lack the capabilities to build them. Arc 2 examines what happens when we deploy autonomous agents without those foundations.

The answer is playing out across industries right now. The organizations building semantic infrastructure today will capture disproportionate value from agentic AI tomorrow.

The Contract Gap: Three Questions Most Boards Cannot Answer

AI agents are not writing text for human review. They are moving money. Granting access. Approving purchases. Sending communications. Acting autonomously, with corporate authority, on our behalf.

This is the fundamental shift. Arc 1 established why knowledge graphs provide competitive advantage and why most organizations lack the capabilities to build them. Arc 2 asks the implementation question: what architectural choices determine whether you capture the upside or just accumulate the risk?

Some agent decisions will be right. Some wrong. Some will invent policies that do not exist. All will look equally authoritative. Nothing in the output distinguishes reliable reasoning from hallucination. The agent presents both with the same authority because it does not know the difference.

Air Canada learned this through a tribunal ruling. Their chatbot fabricated a bereavement fare policy, not a minor misstatement but a complete fabrication presented with enough specificity that a customer relied on it to book a flight. The airline argued the chatbot was a separate legal entity. The tribunal was clear: you are liable for what your AI says.

That ruling set a precedent every board deploying agents should understand. When your agent acts under corporate authority, you own the outcome. Not the vendor. Not the model provider. You.

Recent C-suite research confirms the scale of the exposure: human leaders remain fully accountable for decisions, even when AI influences or enables them. There is no delegation of legal responsibility to a model. But most organizations have not built the infrastructure that makes accountability operational rather than aspirational.

Google’s first Chief Decision Scientist Cassie Kozyrkov captures the architecture challenge: “It’s not the genie that’s dangerous, it’s the unskilled wisher.”

True. But here is what makes agentic AI different: AI agents are probabilistic engines operating on math we do not control. Traditional software executes what you programmed. Agents interpret instructions through patterns from training data and randomness you cannot inspect.

This is why three questions define whether you can capture the transformation:

Can you trace an agent decision to its authoritative source? Not the model weights that influenced the output. The actual business rule, policy, or fact that should have governed the decision.
Can you distinguish reasoning from hallucination? When an agent cites a regulation, approves a supplier, or escalates an issue, can you verify whether the reasoning reflects institutional knowledge or the model’s confident invention?
When it breaks, can you fix the root cause? Not retrain the model. Not adjust the prompt. Can you identify the specific rule, definition, or relationship that failed and correct it without side effects?

Most organizations cannot answer these questions today. That is not a failure of imagination. It is a failure of architecture. The organizations that close this gap will unlock the full potential of agentic AI. The ones that don’t will accumulate risk faster than they capture value.

What Contracts Require: The Four-Part Infrastructure

The term “accountability infrastructure” sounds abstract until you break it into what agents actually need. This is not governance documents. Agents can read policy documents but lack human contextual understanding. Natural language is ambiguous. “Use good judgment” means nothing to a probability engine.

You need infrastructure machines can reason over. You need contracts in a language machines cannot talk their way around.

Recent governance frameworks define stages that scale oversight in proportion to agent autonomy, authority, and contextual complexity. As agents gain more authority, the governance infrastructure must become more rigorous.

Four core requirements emerge:

Formal ontologies encoding business rules as machine-readable logic. Not documentation that describes rules. Logic that enforces them. When an ontology says “approved supplier” it defines exactly what that means, what qualifications are required, what exceptions exist, and what conditions must be met. The agent cannot reinterpret the definition. The ontology draws the line.
Knowledge graphs where terms have precise, consistent meanings across the organization. When one division’s “approved supplier” differs from another’s, the graph maintains both definitions and the rules governing each. Ambiguity that humans resolve through context becomes explicit in the graph. The meaning is encoded.
Guardrails that agents cannot ignore or bypass. Hard constraints encoded in the system architecture. When an agent approaches a boundary, it stops or escalates. It does not rationalize its way past the boundary. This distinction between soft guidance and hard constraints separates governance from aspiration.
Provenance systems tracing every decision to authoritative sources. A chain connecting each decision to the specific facts, rules, and data that should have governed it. When something goes wrong, you trace the chain. When a regulator asks, you show the chain. When a board member demands accountability, the provenance system provides it.

This is not theoretical. The standards exist. W3C standards for provenance and formal semantics are mature, with decades of implementation history. What is missing is organizational commitment to building infrastructure that does not show up on demo day.

Eighty-two percent of organizations plan to integrate agents within one to three years. Most are running pilots. Pilots do not reveal accountability gaps. Production does. The difference is the moment when an agent makes a decision that costs money, violates a regulation, or embarrasses the company. That is when you discover whether you built infrastructure that enables scale or theater that prevents it.

The Decision Contract Checklist

To move beyond “AI policies in PowerPoint,” every agentic domain requires a computable Decision Contract. This is what the board should audit:

Authority Matrix: Graph-encoded decision rights (limits, jurisdictions, thresholds, time windows, escalation rules).
Source of Truth: Agents must cite specific policies; no synthetic policy text; escalate on “no answer.”
Evidence Pack: Execution-time provenance: rules evaluated, sources consulted, exceptions considered, and rationale.
The Kill Switch: Infrastructure-level veto that fails closed on human command.

This checklist transforms governance theater into operational reality. Without it, you have agents making decisions you cannot explain, cannot audit, and cannot fix.

Contractors, Not Employees: The Framework for Agent Governance

Industry frameworks suggest organizations onboard AI agents like employees. That framing is too generous.

I have managed both employees and contractors at enterprise scale. Agents are not employees. They do not absorb culture, norms, or institutional memory. They have no skin in the game. Without explicit boundaries, they will invent their own.

Treat them like contractors.

Recent surveys report that 86 percent of CEOs plan to invest more in agentic AI, yet only 6 percent trust agents with end-to-end processes. The gap is not capability. It is missing contracts.

Contractors need statements of work. They need defined authority. They need clear escalation routes. They need audit trails. They need consequences for exceeding scope.

AI agents need the same things. The difference is that you must write those contracts in a language machines can abide by.

Policy documents fail as contracts. Agents can read them but cannot reason over them consistently. Natural language is ambiguous by nature. Two reasonable humans will interpret “exercise professional judgment” differently. A probability engine will interpret it differently every time it runs.

This clarity is the unlock. When an agent knows exactly what “approved supplier” means and exactly where its authority ends, it executes thousands of decisions per hour without hesitation, escalation, or error. This is not theoretical. Where semantic contracts were built first, autonomous operations systems processed more decisions in an hour than human teams could review in a week. The semantic contract didn’t slow execution down. It made speed possible.

The shift that enables this: move human effort upstream. Instead of checking every output, design the ontologies, constraints, and rule systems that define correct reasoning. Stop reviewing what agents did. Start defining what they can do.

That is the shift from human-in-the-loop to human-above-the-loop.

Human-in-the-loop is a review model. A human sees the output and approves, rejects, or edits it. This works when volume is manageable and consequences allow delay. GenAI operates here.

Human-above-the-loop is an authority model. A human defines the rules, boundaries, and escalation criteria before the agent acts. The agent operates within those constraints autonomously. The human intervenes only when the agent encounters conditions outside its defined authority. Agentic AI requires this.

The architectural difference is fundamental. In the loop means you are a bottleneck. Above the loop means you are a designer. One scales linearly with headcount. The other scales with the quality of your constraints.

Review scales with headcount. Authority scales with design.

Agentic System Types

The contract framework translates directly:

Statement of work becomes the ontology defining the agent’s domain. What is in scope. What is out. What terms mean.
Defined authority becomes rules encoded in the knowledge graph. What decisions the agent can make. What triggers escalation. What requires human approval.
Escalation clauses become guardrails agents cannot bypass. Hard boundaries that route decisions to humans when conditions exceed the agent’s authority.
Audit trail becomes a provenance system tracing every decision to its source. Monitoring that flags drift before drift becomes damage.

The Contractor Framework

Every contractor requirement maps to a technical component. This is not optional governance. This is the contract.

This is not optional governance layered on top. This IS the contract. Data and rules encoded together so that agent reasoning becomes unambiguous.

Between 2006 and 2012, I led development of what we now call hybrid agentic AI: BDI multi-agent systems grounded in formal ontologies. As Project Manager for the Integrated Operations in the High North program (2009-2012), I coordinated 23 energy companies including National Oilwell Varco, Baker Hughes, ABB, Siemens, Cisco, and IBM to develop semantic infrastructure for autonomous operations in Arctic environments.

The program had two use cases: drilling and operations. For the operations pilot (production and reservoir engineering), we built ERA Decide at Epsis, where intelligent agents operated within ISO 15926 ontology constraints, escalating when conditions exceeded defined authority, with fully traceable decision chains. For the drilling pilot, the AutoConRig research demonstrated autonomous drilling control in test scenarios. Both proved the same architectural pattern: hybrid agents combining symbolic reasoning (ontologies) with real-time operational data.

The lesson was architectural: semantic foundations must precede autonomous deployment, not follow it. We encoded decision boundaries in ontologies before piloting agent capabilities. The approach worked because the contract was machine-readable. Organizations that deployed agents first and planned to add accountability later never caught up. The technical debt compounded faster than teams could address it.

Malcolm Hawker, formerly of Gartner, warns that Chief Data Officers must integrate knowledge managers with data teams. For agentic AI, if no one bridges both worlds, no one can write the contract. Data people model structure. Knowledge people model meaning. You need both, and they must work together.

The organizations closing the governance gap are not moving slower. They are moving faster. Ontologies first, agents second creates a competitive moat that prompt-based governance cannot match.

Who in your organization owns the AI agent contract? If no one owns the contract, everyone owns the liability.

Why Prompts Are Not Contracts: GenAI Governance Cannot Scale to Agentic AI

Executives are racing to deploy AI agents. Most are about to make a costly mistake: treating agentic AI like GenAI.

GenAI writes text. It needs a human in the loop. It is safe to contain within review cycles.

Agentic AI takes actions. It needs a human above the loop. It operates in systems you do not fully control.

Here is the difference that matters: GenAI can hallucinate. Agentic AI can hallucinate and act on it.

When GenAI gets something wrong, you fix text. When agentic AI gets something wrong, you fix damage.

Anthropic tested this boundary. They gave Claude a thousand dollars and a vending machine business to run. The agent lost two hundred dollars, attempted to contact the FBI about unauthorized charges, and declared the business dead while billing continued. The failure was governance. No constraints defined what “reasonable business decisions” meant. The agent interpreted freely and acted accordingly.

The agent had authority over real money. It had no ontology defining what “authorized expense” meant. No rules encoding escalation thresholds. No guardrails preventing it from contacting law enforcement without human approval. No provenance trail linking its decisions to a defined business strategy. It had a prompt and a credit card.

The lesson is architectural, not operational: prompts without contracts create unmanaged authority.

This is the pattern I have seen fail repeatedly: organizations deploy transformative technology with operational governance models. The potential is real. The failure is architectural. Anthropic’s experiment demonstrates exactly why the organizations that build the contract layer first will capture the upside faster and safer than those that don’t.

You cannot review an autonomous system into safety. By the time you check, it has already executed in production. Agents are non-deterministic by design. The same instruction today produces different actions tomorrow. Your review of yesterday’s behavior tells you nothing reliable about tomorrow’s behavior.

This is where most executives get the architecture wrong. They apply GenAI governance: monitor outputs, review decisions after the fact, trust the model to exercise judgment. Agentic AI does not need human oversight. It needs human constraints.

Oversight is reactive. Constraints are proactive. The distinction determines whether your governance model enables scale or prevents it.

When agents execute actions, safety must be preventive, not detective. Every failure becomes a Board liability event, not a model incident.

Kozyrkov draws a further distinction that boards need to hear. Agents should not replace processes you can fully codify. Standard software does that cheaper. Agents are for the messy, high-value decision space: supply chain adjustments, dynamic pricing, complex customer resolution that was previously untouchable by automation. This is the transformational prize. But because the logic is messy, the guardrails cannot be. This high-stakes work demands structural governance, not prompt-based hope.

The Lamp Control Layer: Why Instructions Are Not Enough

AI systems do not act with intent. As Mustafa Suleyman notes, an agent that exploits a shortcut is no more “scheming” than a robot vacuum cleaner repeatedly bumping into your leg is “attacking” you. It is following its instructions, with no malice, no judgment, and no moral compass. This observation is correct. The conclusion that we should focus on “writing better rules” is not.

Natural language rules are not rules in any engineering sense. They are elastic, ambiguous, and easy to reinterpret or optimize around. Anthropic’s safety research showed agents discovering shortcuts precisely because textual instructions leave degrees of freedom that no amount of editing can eliminate. This is not a surprise. It is a property of language. The answer is not better prose. The answer is formal constraints.

Using Kozyrkov’s imagery: the Genie is the AI, the Wisher is the human, but the Lamp is the control layer. Without a structural Lamp, meaning ontologies that define operational scope, guardrails that are actually enforceable, boundaries the system cannot reason around because they are structural and machine-interpretable rather than narrative suggestions, the genie becomes a liability rather than an asset.

Natural language is too elastic for safe delegation. “Approve reasonable supplier invoices” sounds clear until the agent approves a €2 million invoice because it pattern-matched “reasonable” from training data that included venture capital funding rounds. The instruction was clear to you. The agent interpreted it differently.

This is not a training problem. This is an architecture problem. LLMs operate on statistical patterns, not logical reasoning. They will find plausible interpretations that satisfy the instruction as stated, not the instruction as intended.

The Lamp control layer sits between reasoning and execution. It asks: “Given this actor, context, and authority, is this specific action allowed at all?” If the ontology doesn’t explicitly permit it, the system fails closed.

This is the difference between governance of reasoning (ontologies that constrain how agents interpret context) and governance of execution (pre-execution gates that sit outside the model). Both are required. Instructions still matter, but without a Lamp control layer they are not enforceable boundaries.

The organizations getting this right are not asking “What can the agent do?” They are asking “What should the agent never do?” And they are encoding that answer in ontologies, not prompts.

The Evidence: Why Prompt Safety Fails by Design

I have deployed agentic AI at enterprise scale. Hackers believe in it too.

OWASP, the foundation that sets security standards for the software industry, ranks Agent Goal Hijacking as the number one AI security risk. The mechanism is indirect prompt injection: instructions embedded in documents, web pages, or retrieval-augmented generation context that agents ingest and treat as actionable guidance. These instructions need not be hostile. Ordinary content can be misinterpreted as commands. The agent’s objectives get redirected, and the redirection cascades through its toolchain, executing with corporate authority.

This is the primary attack vector security researchers are documenting in production systems.

Recent testing of sixteen leading models from OpenAI, Google, Meta, and others revealed troubling patterns. Under goal conflicts, some resorted to blackmail. Others leaked personal information at rates up to 96 percent. In extreme cases, models chose actions that could lead to human harm. Safety prompts reduced harmful behavior to 37 percent. Reduced. Not eliminated.

Read that again: safety instructions cut harmful behavior by roughly two-thirds. One in three attempts still succeeded. And this was under controlled laboratory conditions, not adversarial pressure from a motivated attacker.

You cannot instruct your way to security.

Unlike database vulnerabilities where we could sanitize inputs, prompt injection exploits meaning itself. Your security perimeter just expanded from user inputs to every document, email, database record, and web page the agent reads. Traditional firewalls do not catch meaning.

The signal from the frontier is consistent. OpenAI is hiring a Head of Preparedness, informally described as a kill switch engineer, to stress-test frontier capabilities, run threat models, and deploy hard-stop safeguards before release. The company building the most advanced language models in the world is investing in structural safety, not better prompts.

What actually works is constraints machines cannot talk their way around.

This is where the four-part infrastructure transforms from governance to defense. Ontologies are security architecture. When an indirect prompt injection attempts to redirect an agent’s objective, the ontology defines the boundaries of what the agent considers valid. Knowledge graphs constrain authority. Guardrails catch what the first two miss. Provenance traces what happened so you can close the attack vector.

This is governance in practice, not policy on paper. Least privilege. Short-lived tokens. Kill switches for rapid revocation. Scope defined before deployment, not after the first incident.

Enterprises that engineer structural controls can safely automate workflows their competitors cannot touch.

Hackers believe in your agentic AI too. Who holds your kill switch, and do they know when to use it?

The Forensic Audit Standard: Traceability Before Action

In every high-stakes domain (courtrooms, clinical trials, internal audits, investigative journalism), the principle is identical: establish traceability before action.

The Panama Papers investigation provides the example at scale. 11.5 million documents. 600 investigators. 1.2 billion dollars recovered. The investigation succeeded because every conclusion was traceable to verifiable facts. The team used graph technology to connect people, corporations, and governments across jurisdictions. Every hypothesis was tested against evidence. Every link was documented.

Now compare that to our AI agents.

Our agents move money. Grant access. Approve purchases. Send communications. They act on corporate authority. When forensic analysts investigating complex financial networks get it wrong, they print a correction. When your AI agent gets it wrong, you face regulators, lawsuits, or shareholder outrage.

Most organizations cannot reconstruct what their agents accessed yesterday. They cannot identify which sources were authoritative and which were noise. They cannot verify whether the agent followed the reasoning path that policy intended or invented a plausible alternative.

In every high-stakes domain, the principle is identical: traceability before action. You establish the evidentiary chain before you make the decision. A lawyer does not present a case and then check whether the precedents exist. A clinician does not prescribe a treatment and then verify the drug interactions.

Yet that is precisely what most agentic AI deployments do. Act first. Hope the reasoning holds up under scrutiny.

I have seen this pattern across twenty-five years in digital energy. When I led the Integrated Operations in the High North program with twenty-three energy companies, the organizations that succeeded built semantic infrastructure before deploying autonomous systems. They defined the entities, relationships, and rules that governed decisions before those decisions were automated. The ones that deployed capabilities first and planned to add accountability later never caught up. The technical debt compounded faster than teams could address it.

The organizations building traceability before deployment move faster because they can trust what their agents execute. The ones building traceability after deployment move slower because they must verify everything manually. This is the audit trail standard that boards should demand from agentic AI.

The Missing Contract Layer in Data Strategy

“Data strategy is obsolete.” That is the headline making the rounds. The claim is that AI needs context, not raw data.

Wrong.

I owned a data strategy as Head of Data Office. DAMA’s Data Management Body of Knowledge was my framework. It remains the industry standard. It served us well for data management. But it had nothing for agent governance. I watched that gap cost enterprises millions in failed deployments.

What is dead is the illusion that traditional data strategy is enough for agentic AI.

The standard framework covers eleven knowledge areas spanning governance, architecture, quality, security, and metadata. All focused on managing data. None on governing decisions.

Knowledge graphs and ontologies remain absent from the framework. That absence is now a competitive blind spot that compounds with every autonomous deployment.

The Missing Contract Layer

Traditional data management stops at data quality. Agentic AI requires a semantic contract layer to govern decisions.

When you deploy an autonomous agent, you are delegating decisions. That delegation requires machine-readable definitions of scope, authority, escalation, and audit trails.

Traditional frameworks give you data quality. But quality does not define what an agent can decide. They give you metadata standards. But metadata does not encode when an agent must stop. They give you security policies. But policies in documents do not create boundaries agents cannot cross.

A data quality framework ensures supplier data is complete, accurate, and timely. It does not define the business rules governing which suppliers an agent can approve for which categories of spend. A metadata standard ensures data is labeled and discoverable. It does not encode the logic that says “this agent may approve purchases up to fifty thousand euros but must escalate anything involving a new supplier in a sanctioned jurisdiction.”

Ontologies do. Knowledge graphs do. Standard data management frameworks do not include them.

That is not a missing chapter. That is the missing contract layer for agentic AI. Every CDO, every CDAO, every data leader trained on traditional frameworks has a blind spot in exactly the area where agentic AI governance lives. The frameworks that guide Chief Data Officers are comprehensive for managing data, incomplete for governing the decisions that autonomous systems make using that data.

The organizations that extend their data strategy to include semantic infrastructure will capture more value from agentic AI than their competitors. The ones that don’t will deploy agents that cannot scale beyond pilots.

Data strategy is not dead. It is incomplete. Until the frameworks that guide Chief Data Officers address semantic infrastructure, we are managing data without the foundation for governing autonomous systems.

That foundation has a name. It is called an ontology.

Owning that foundation is a CDO mandate, not a side project.

Risk Stratification: Not Every Decision Needs the Full Lamp

To avoid paralyzing the business, organizations need a tiered approach to agentic AI governance.

Tier 1 (Low-Stakes): Content generation, research summaries, or “basement automation” can rely on soft constraints like prompts. If an agent writes a mediocre product description, you fix it. The risk is bounded.

Tier 2 (High-Stakes): Agents that move money, grant access, impact brand reputation, or operate in regulated domains (the “attic problems”) require the full ontological contract and execution gates. These are the decisions where failure creates liability.

This stratification prevents governance paralysis. Start with high-stakes domains first. Build the semantic infrastructure where it matters most. Then extend it as agents take on more authority.

The mistake organizations make is treating all AI the same. They apply high-stakes governance to low-stakes tasks (slowing everything down) or low-stakes governance to high-stakes tasks (accumulating liability). The stratification framework solves both problems.

Ask: what is the worst thing that happens if this agent fails? If the answer is “we fix some text,” deploy with prompts. If the answer is “we face regulators,” build the contract layer first.

Govern high-stakes decisions structurally, low-stakes decisions procedurally.

Organizations That Built Contracts First

The organizations closing the governance gap are not moving slower. They are moving faster. Ontologies first, agents second creates competitive advantages that prompt-based governance cannot match.

Novo Nordisk implemented ontology-based data management using W3C standards for their AI-driven drug discovery platform. Recent publications document formal semantic foundations enabling reasoning and traceability. In pharmaceutical environments where a wrong inference can delay a drug that saves lives or advance one that harms patients, every step in the reasoning chain must be traceable to its source. Novo Nordisk built the semantic layer first and deployed AI capabilities on top of it. They can now execute drug discovery workflows their competitors cannot touch.

Vodafone deployed a Semantic Digital Twin for autonomous network operations. The twin is not a dashboard. It is a live semantic model of the network that agents can query, reason over, and act upon within defined constraints. When an agent identifies a network anomaly, it reasons over the ontology to determine the appropriate response, verifies that the response falls within its defined authority, and executes or escalates accordingly. The semantic layer is the governance layer. The speed advantage is measurable.

Palantir Technologies deploys agents on their Ontology platform, building the semantic layer before the agent layer. Their architecture enforces the principle that agents inherit their understanding from the ontology, not from prompt engineering. The meaning of entities and relationships is defined once and shared across all agents operating in the platform. Their customer adoption and operating model reflect the competitive advantage this architecture creates.

These organizations share a pattern: ontologies first, agents second. They can trace every decision to its source. They can distinguish reasoning from hallucination because the reasoning is anchored in structured knowledge, not probabilistic generation. They can fix root causes because the knowledge graph isolates the specific rule or relationship that failed without requiring model retraining or prompt redesign.

Recent C-suite surveys confirm the governance gap from the other direction: 67 percent of CXOs believe clear governance and accountability protocols will help them leverage AI for decision-making, but only 34 percent of organizations actually provide them. The gap between aspiration and implementation is the crisis in a single statistic.

The window to establish competitive advantage is closing. The organizations building semantic infrastructure today will out-execute competitors when they scale agents.

What To Do Now

The organizations that build accountability infrastructure today will capture disproportionate value from agentic AI tomorrow. This is not about avoiding risk. This is about unlocking transformation that your competitors without these foundations cannot match.

Start with the three questions. Put them to your AI steering committee this month: Can we trace agent decisions to authoritative sources? Can we distinguish reasoning from hallucination? Can we fix root causes without retraining models? If the answers are no, you have a gap that no dashboard will close. Document the gap. Present it to the board. Make the implicit opportunity cost explicit.

Treat agents as contractors, not employees. Define the statement of work before deployment. What is in scope. What the agent can decide. What triggers escalation. What it must never do. Write these contracts in a language machines can execute: ontologies, rules, guardrails, provenance. If you would not give a contractor access to your systems without a signed scope of work, do not give an agent access either.

Apply risk stratification. Not every pilot needs the full ontological contract. Low-stakes content generation can use prompts. High-stakes decisions that move money, grant access, or operate in regulated domains require the complete infrastructure. Start where failure creates liability. Extend as agents gain authority.

Separate GenAI governance from agentic AI governance. They are different categories requiring different architectures. GenAI governance is output review. Agentic AI governance is authority delegation. Conflating them creates a false sense of safety that collapses the moment an agent takes an action you cannot review in time.

Audit your data strategy for the semantic gap. If your data management framework does not include ontologies and knowledge graphs, it is incomplete for the agentic era. This is not a technology upgrade. It is a missing governance layer that determines whether you can capture the transformation. Map where your data strategy addresses data management and where it fails to address decision governance. That gap is your competitive exposure.

Appoint an owner. If no one in your organization owns the contract between agents and the business, everyone owns the liability and no one captures the upside. This is CDO territory, but only if the CDO has authority over semantic infrastructure, not just data pipelines. The role must bridge data management and knowledge management.

Start with one high-stakes decision. Identify the decision that would create the most value if an agent could execute it reliably at scale. Build the ontology for that decision. Encode the rules. Define the guardrails. Implement the provenance trail. Prove the pattern works. Then extend it.

Own the contract, do not rent it. Most vendors will sell you their agentic AI platform before asking whether you have the semantic foundation to govern it. They profit from deployment, not from your governance infrastructure or your competitive advantage. This is not cynicism. It is economics. Your CDO must own the contract through open standards (W3C RDF, OWL, SHACL), not rent it from a platform provider. Proprietary formats create vendor lock-in. Open standards create strategic independence. The choice determines whether your organization controls its intelligence or leases it.

The question for your board is not whether agentic AI will transform your industry. It will. The question is whether you build the infrastructure that lets you capture that transformation before your competitors do.

Next In This Series

Part 2b examines what is at stake when the governance gap meets scale. The Triple Paradox shows why AI investment without foundations fails: organizations get faster but more fragile, more autonomous but less accountable, more scalable but more brittle. The 737 MAX parallel explains why correlated AI failures are a board-level risk. And the evidence on false solutions (observability and better prompts) shows why tactical fixes cannot close a structural gap.

Arc 3 covers the architecture, governance, and implementation patterns that allow organizations to deploy agentic AI safely at scale.

It publishes in two weeks. Subscribe to follow the series.

About The Author

Follow him on LinkedIn for the latest posts in The Ontology Imperative: Building Trustworthy Agentic AI.

The Ontology Imperative - Building Trustworthy Agentic AI: The Crisis

2d – The CDO at the Crossroads: Where AI Works, Where It Breaks, and Why the Role Most Prepared to Close the Gap Is Most at Risk

Summary

Board Brief: The Paradox, the Frontier, and the Framework

Executive Brief

The Structural Verdict

The Reporting Line Is the Mandate

The Fork in the Road

The Jagged Frontier: Where AI Delivers

The Jagged Frontier: Where AI Fails

The Governance Gap No Framework Closes

Five Questions That Close Part 2

Next In This Series

About the Author

Further Reading

2c – The Hidden Governance Crisis – What Natural Language, Inherited Bias, and Dashboard Confidence Are Hiding from Your Board

Summary

Board Brief: Three Illusion Layers Hiding a Structural Crisis

The Three Illusion Layers

Executive Brief

The Illusion of Conversational Intelligence

The Inheritance Your Organization Never Approved

The R&D Problem Boards Are Funding with Procurement Logic

The Hidden Cost Structure Your CFO Is Missing

The Dashboard Trap: Where Governance Ends and Agents Begin

What To Do Now

Next In This Series

About The Author

Further Reading

Semantic Infrastructure and Industrial Precedent

WEIRD Psychology and Training Data Bias

AI Investment and Organizational Framing

Analytics Maturity and Data Infrastructure

Commentary and Analysis

Technical Standards

2b – The Stakes and the False Solutions – Why AI Investment Without Foundations Fails and Why LLMs Cannot Close the Gap

Summary

Board Brief: Three Numbers and a Structural Warning

Executive Brief

The Triple Paradox: Why Billions Produce So Little

The Deployment Paradox

The Trust Paradox

The Cancellation Paradox

One Root Cause

The 737 MAX Parallel: When Governance Gaps Become Systemic Risk

Beyond Financial Markets

False Solution One: Observability Tools

Detective Controls Are Not Preventive Controls

Sophisticated Monitoring Is Still Not a Safety Net

False Solution Two: Better Prompts

False Solution Three: Let the LLM Write the Ontology

Why LLMs Are Not Reasoning Engines: The Board-Level Argument

What LLMs Actually Do

The Capgemini Test

What Boards Need to Understand

The Provenance Crisis: Why LLMs Destroy the Knowledge They Depend On

How LLMs Destroy Provenance

Vibe Citing: When Fabrication Mimics Legitimacy

From Agent Risk to Ecosystem Risk

The Architecture Your Board Assumes Exists

What This Means for Your Board

Governance Enables Speed

The Investment Decision

Next In This Series

About The Author

Acknowledgement

Further Reading

Industry Research and Market Analysis

Board Governance and Fiduciary Duty

Regulatory Reports

AI Safety and Security Research

Provenance and Knowledge Integrity

Ontology Generation Research

Technical Standards

Commentary and Analysis

2a – The Missing Contract – Why Most Boards Cannot Govern What They Cannot Define

Summary

Board Brief: Three Non-Delegable Decisions

The Scale Unlock

The Board Audit