AI Governance: The Number Gets Worse Before It Gets Better
Sixty-four per cent of enterprises have traced a confidently wrong AI agent answer to a problem in their own company data. That happened within the past six months [1]. That is the number most of this week's coverage led with. The one underneath it should worry a board more. It runs backwards from what anyone expects to hear.
Split that sixty-four per cent group one more way: by whether the organisation has a governed semantic layer. That is the system meant to tell an AI agent which definition of revenue is current, which table is stale, and which number the business actually agreed on. Split that way, the picture inverts. Among organisations running, piloting or building one of these layers, seventy-eight per cent report a context failure. Among organisations only evaluating one, or with no plans at all, the figure is thirty-seven per cent [1]. The group that built the control reports more than twice the failure rate of the group that did not.
VentureBeat's own researchers surveyed a hundred and thirty qualified respondents, drawn from their own reader and panel base rather than a probability sample. They say plainly that they did not test why the two groups differ. The honest reading of seventy-eight against thirty-seven depends entirely on that gap.
Two explanations sit underneath that comparison. The survey cannot tell them apart. The first is that building a governed semantic layer somehow introduces new failure modes of its own. The second is that building one finally lets an organisation detect failures its agents were already producing, quietly, against ungoverned context, for however long those agents have been running. A board that has not been shown both readings side by side will reach for the first one. It is the simpler story. It fits an expensive new system being blamed for a worse-looking result. The macro data below backs the second reading.
A governed semantic layer works like a glossary with teeth. It is a single place where "revenue," "active customer" and "current quarter" mean one thing, enforced automatically rather than argued over in a meeting every quarter. Enterprises have been building something like this for longer than generative AI has existed. Two finance teams quoting different numbers for the same metric is an old problem with an old cost. What changed is that an AI agent now reads from whichever system answers fastest. A glossary that only humans consult does nothing to stop that.
VentureBeat's tracker, built from monthly waves of its own VB Pulse research panel, has been asking enterprises about exactly this gap since at least June. Read in isolation, the seventy-eight against thirty-seven comparison looks like an indictment: the expensive control made things worse. Read against what the organisations without a layer cannot see, it looks like something closer to a measurement effect. An organisation with no inventory of its own context problems has nothing to report. Nobody has built anything capable of finding them yet.
The naive reading says a governed semantic layer causes failures. What it has actually done is put a measuring instrument where there was previously only guesswork. A sharper instrument finds more faults, the way a sharper torch finds more cracks in a wall that has had them for years. The fault was always there. Seventy-eight per cent measures how often governance now notices a problem that unmeasured operations were already living with. It is not how often governance creates one.
This matters because of a common instinct in the boardroom. A first-year assurance programme's numbers look worse than last year's silence, and the instinct is to treat the programme as the problem. This is the conversation every assurance lead eventually has to have. This year's honest count looks alarming next to last year's absence of one. The absence was the worse of the two, not the safer one.
None of this excuses the underlying thirty-seven per cent either. An organisation that has not built a semantic layer and reports fewer context failures has not solved the problem. It has simply lost the ability to see it. Seventy-eight per cent measured and thirty-seven per cent unmeasured is not a fair comparison. Treating it as one, in either direction, misreads what the two numbers are actually counting.
The finding that sharpens this argument, rather than undermining it, sits one layer further in. Among the organisations running, piloting or building a governed semantic layer, only thirteen per cent say that layer is where their AI agents actually get business context. That figure is from the most recent wave. Nineteen per cent said so the month before that, and twenty-one per cent the month before that [2]. Most of the rest draw on something else entirely: retrieval over whatever documents happen to be indexed, direct queries to whichever live system is reachable, or long-context loading that pulls in everything. The model sorts the rest out itself.
This finding points the same way. The organisation built a control, and its own measurement found a context problem. It still has not closed the loop between that control and the agents it runs day to day. A governed definition that nothing consults is only ever documentation, however carefully it was written and however well funded the project behind it.
Readers of this series' Wednesday instalment will recognise the shape. The Linux kernel's maintainers added a public AGENTS.md convention telling automated contributors how a change should be reviewed before it is accepted. Whether that convention binds anything depends entirely on whether the agent submitting the change bothers to read it. Two layers of the same technology stack, a kernel review process and an enterprise context layer, have arrived at the same fact. A rule has to be consulted, not just written down. Writing it down is the easier half of the job.
The clearest image here is a safety inspector walking into a machine room for the first time with a proper torch. The floor has always had only a bare bulb. The cracks the torch finds were not put there by the inspection. Ordinary wear put them there years ago. Nobody had the light to see them before now. The inspector who reports twelve new faults has not made the floor more dangerous. The inspector who reports nothing, because the bulb never reached the far corner, has not made it safer.
The discipline this calls for is a clear separation between the rate at which a fault is found and the rate at which a fault exists. That separation belongs in every report that follows, not fewer inspections. Those two numbers move at different speeds in the first year or two of any proper measurement programme. Conflating them is where most of the boardroom alarm comes from.
The same pattern shows up one level higher in the stack, in the macro figures McKinsey published in August. Enterprise-wide AI scaling rose from thirty-eight to forty-four per cent year on year. Over the same period, the share of organisations reporting any positive EBIT contribution from AI stayed flat at thirty-seven per cent. The share McKinsey classifies as AI "high performers" held at roughly six per cent [3]. McKinsey defines that tier as attributing at least five per cent of EBIT to AI with a self-described significant effect.
Forty-four per cent moved six points in a year. The return figure moved none at all. Deployment scaled. Measured return did not. Two explanations fit that gap. Assurance and governance programmes, in their first year or two of honest measurement, sit in the phase where they make the gap between deployment and value visible, rather than close it. A chief information officer who has just finished building the instrumentation that finally shows this gap clearly is looking at the first accurate year her organisation has had. It is not a worse one.
That reading does not let vendors or boards off the hook for the underlying thirty-seven per cent. It explains why the number has stayed flat while adoption kept climbing. That is not proof the whole exercise has failed. An organisation that mistakes this year's honest measurement for this year's deterioration will do the one thing that actually makes deterioration likely. It will defund the layer that was starting to tell it the truth.
The "high performers" figure has barely moved while everything around it has. Roughly six per cent of organisations attribute at least five per cent of EBIT to AI with a self-described significant effect. That share has held near flat even as scaling climbed six points in a year. If governance maturity were simply catching up to deployment on a predictable lag, a reasonable board would expect that top tier to widen a little as more organisations clear the bar. It has not widened. That is consistent with a slower, structural story rather than a short lag. Real value appears to go to organisations that pair deployment with the measurement to know whether it worked. That pairing is still rare, six points of fresh scaling notwithstanding.
No New Zealand-specific survey measures this exact pattern, context failure against semantic-layer adoption. A loosely related domestic figure would add nothing to a global finding that already makes its point.
This calls for reporting what was actually measured, in a shape a board can act on rather than panic over. It does not call for measuring less. Four things belong in that report before the headline number does.
Measure before you judge. A failure rate reported in isolation, with no baseline for how many failures existed before anyone was counting, invites exactly the wrong comparison. It compares this year's honest seventy-eight against an imagined zero from the year nobody looked. The honest comparison is this year's seventy-eight against an estimate of what was actually happening last year. That estimate can be rough at first.
Separate detection from rate. A programme's first annual report should state, in plain words, how much of this year's increase is new instrumentation finding old problems. It should also state how much is a genuine change in how often the system fails. Those are different numbers. They call for different responses. Collapsing them into one line is how a working control gets read as a broken one.
Check what the agents are actually reading. Thirteen per cent is most of the point, far too large to wave off as a rounding error. Before presenting a semantic layer as a solved problem, confirm the agents in production are actually calling it. Otherwise they may be retrieving from whichever document index happened to be reachable first. A control an agent does not consult is a line item rather than a safeguard. The gap between the two does not show up on an architecture diagram.
Report the baseline honestly, even when it is unflattering. A board that hears "we found more problems this year because we finally looked properly" the first time a measurement programme runs will trust the second year's numbers. A board that is handed a clean dashboard this year will see a messier one once real measurement starts. It will not know the difference between a programme that is working and one that is failing. Nobody told it there was a difference to watch for.
These four steps separate a chief information officer who can defend a worsening-looking number in front of a board from one who cannot. Both start from exactly the same underlying facts.
Scale this out a year or two. The organisations that get it right will be the ones that treated the first honest number as a baseline worth defending, not a crisis worth hiding from. The ones that get it wrong will spend that year quietly dismantling the one part of the estate that was finally capable of telling them anything true. The number it produced looked bad in the one meeting that mattered most.
The open-source dimension sits underneath the specific tool most enterprises are reaching for to fix this. The Model Context Protocol, the open specification Anthropic published in November 2024 and handed to the Linux Foundation's Agentic AI Foundation in December 2025 as a vendor-neutral standard governed the same way Kubernetes is [5], is the mechanism by which an agent is actually supposed to reach a governed context source rather than guess at one. An open specification with a published governance board is checkable in a way a vendor's internal integration layer is not: any organisation can read how an agent is meant to request context, rather than trusting a single supplier's account of its own plumbing. The same openness that makes the kernel's contribution rules auditable makes an agent's context path auditable too, and a context layer built on a closed, single-vendor integration carries a sovereignty question underneath the deployment question.
The implication on the sovereignty side is the same pattern running through a government's own evaluators. In the United States, federal agencies are required to maintain a public inventory of their AI use cases, and the Government Accountability Office's own audit of that requirement found that most agencies' inventories were incomplete or inaccurate once GAO actually checked them against the standard [4]. An agency that built the inventory is, by definition, the one whose gaps GAO's reviewers can find; an agency that never built one reports nothing wrong, because nobody has looked, not because there is less to find. A sovereign evaluator inherits the identical detection penalty a chief information officer does. The audit trail has to exist before the gap in it can be counted, and building that trail is the harder, less flattering half of governance, in Washington or in any enterprise reading this on a Tuesday.
Has your organisation's first honest measurement of an AI system ever looked worse than the silence it replaced? Did the room read that as the problem, or as the proof the problem had finally been found?
If your organisation is moving AI agents from pilot to production and nobody outside the vendor has inspected the control plane, message me and I will send the scope and the fixed fee for an independent review.
• • •
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. I write as director of Te Pono Limited; the views are personal and do not represent the position of any client, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance, and it is politically neutral.
• • •
Andreas Hamberger is a New Zealand leader in Architecture & Security and Associate Member of the Institute of Directors. The Hamberger Report: Generative AI 2026 provides enterprise leaders with evidence-based analysis of the AI landscape. Through Te Pono he provides independent reviews of agentic AI control planes for organisations moving from pilot to production; contact andreas@thehambergerreport.com for the scope and fixed fee.
This article was produced with AI assistance under my direction. Research, drafting and images pass through a pipeline I built and govern: automated gates for source verification, forbidden language and political neutrality, and my own review before anything is published. The tools include Claude, Gemini and Openart. The frameworks, arguments and editorial judgements are mine and are the same discipline I apply to the AI systems I audit for clients. AI accelerated the work; the thinking, and the responsibility for it, are mine.
[1] VentureBeat. "Agentic context layers: 64% of enterprises traced a confidently wrong AI agent answer to problems in their own company data in the past six months." 5 October 2026. https://venturebeat.com/resources/agentic-context-layers-64-of-enterprises-traced-a-confidently-wrong-ai-agent-answer-to-problems-in-their-own-company-data-in-the-past-six-months
[2] VentureBeat. "Companies are building governed semantic layers. Most say their AI agents are getting context somewhere else." 6 October 2026. https://venturebeat.com/data/companies-are-building-governed-semantic-layers-most-say-their-ai-agents-are-getting-context-somewhere-else
[3] McKinsey & Company. "The State of AI in 2026." August 2026. https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi-aug-2026.pdf
[4] United States Government Accountability Office. "Artificial Intelligence: Agencies Have Begun Implementation but Need to Complete Key Requirements." GAO-24-105980. 2024. https://www.gao.gov/products/gao-24-105980
[5] Model Context Protocol. "MCP joins the Agentic AI Foundation." 9 December 2025. https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/