Where the oversight went

Somewhere between the pilot and the production rollout, the oversight went missing.

That is not an accusation. It is a survey result. Dataiku's 2026 Global AI Confessions report — a Harris Poll of 685 CIOs at large enterprises — reads less like a governance benchmark than a quiet admission. 81% of CIOs say they have lost oversight of the AI agents running in their organisations. 83% say there is no standardised lifecycle management for those agents. 75% cannot monitor all of their agents in real time. And 47% have already decommissioned more than twenty agents this year.

87% of the same CIOs say agents are already embedded somewhere in the enterprise. 62% say those agents sit inside business-critical workflows.

Read the numbers together and the shape of the problem appears. The agents are in the building, doing consequential work. The oversight is not.

Oversight was designed for a slower animal

It is tempting to read the confession as negligence. It is closer to arithmetic.

Enterprise oversight mechanisms were built for a workforce that moved at human speed. A new hire gets a manager. A consequential action gets an approval step. A queue of exceptions gets a reviewer, and the queue moves at the rate a human can read. Every one of these mechanisms assumes that the thing being supervised produces work at roughly the rate the supervision layer can absorb it.

Agents break that assumption in both directions. They multiply faster than review capacity — a workflow that produced fifty exceptions a week produces five thousand decisions a day once an agent runs it. And they arrive faster than governance cycles — a team can stand up an agent in an afternoon; the oversight policy for it takes a committee a quarter. The oversight layer did not fail. It was simply never resized.

The result is the configuration the survey describes: agents in production, embedded in critical workflows, with nobody watching who could actually intervene.

Watching is not attending

The instinctive response is a dashboard. If 75% cannot monitor their agents in real time, buy better monitoring. And monitoring has its place — it is how you know an agent is running, what it touched, what it spent.

But a dashboard is not oversight, any more than a security camera is a guard. Monitoring tells you what happened. Oversight is the authority to stop something before it happens, attached to someone with the judgement to know when. A CIO staring at a real-time feed of forty agents is not overseeing them. She is watching forty things she cannot possibly attend to. Universal surveillance collapses into the same skim-approval that blanket review did — attention divided until it is no longer attention.

The distinction that matters is between watching and attending. Watching is continuous, undifferentiated, and nobody's name is on any particular decision. Attending is selective and owned: a defined boundary where consequential actions stop, and a named, credentialed human on the other side of it.

The off switch is not an oversight model

The most telling number in the report is the decommissioning one. Nearly half of the CIOs surveyed have switched off more than twenty agents this year.

Some of that is ordinary technology turnover. But read it alongside the 81% and a quieter interpretation emerges: when oversight cannot be established, the only control left is termination. An agent nobody can supervise does not get supervised. It gets switched off. The off switch becomes the governance model of last resort — crude, total, and expensive in a way that never shows up on a risk register. The workflow goes back to manual, the throughput gain evaporates, and the next agent proposal dies in committee because the last one ended in an uncontrolled deployment.

This is the actual cost of lost oversight. Not the incident that makes the news. The quiet reversion — work that goes back to being done slowly, by scarce senior people, because the organisation never built the structure that would let an agent do it safely.

Oversight is a structure, not a feed

The alternative is not more surveillance. It is a different architecture, and it is not exotic. It looks like the oversight structures enterprises already trust for human work, resized for machine throughput.

A named agent with a defined scope — not a shared service account doing everything for everyone. An explicit consequentiality threshold — which actions may execute on the agent's own judgement, and which stop until a human approves them. A credentialed expert at that boundary — someone whose professional standing is the thing being exercised, not a ticket queue. Logging that can reconstruct any decision: who approved it, when, and on what basis. And a correction loop, so that every expert intervention becomes training signal — the oversight layer improving the thing it oversees, compounding rather than depleting.

This is the attended model h.work is built on, and the survey numbers describe its absence almost exactly. The 81% who lost oversight did not lose it to a sophisticated failure. They lost it to deployment without structure — agents that arrived as tools and never acquired the scaffolding of a worker: a name, a scope, a supervisor, a record.

The question for the next review

The confession reports are useful because they convert a private anxiety into a public baseline. Four out of five enterprise technology leaders now say, on the record, that they cannot account for the agents running in their organisations. That is not a fringe position. It is the median enterprise condition in 2026.

Which reframes the question. The useful one for the next operations review is not "how do we watch the agents" — watching is the thing that already failed. It is: which of these agents' decisions need a name on them. Answer that, and the monitoring problem shrinks to size. The routine work can run unattended, logged and sampled, because someone has decided it is routine. The consequential work stops at a boundary where a credentialed human attends, approves, and owns.

The moat was never intelligence. It is attention — and attention, it turns out, is the first thing most deployments forgot to build.