Only 21% Have Mature Agent Governance. Score Yourself Against the Other 79%

Deloitte surveyed 3,235 business and technology leaders in 24 countries. 21% said their organisation has a mature governance model for autonomous agents. Roughly three quarters plan to run agents at moderate scale or higher by 2027. AI governance maturity is not keeping pace, and most teams cannot tell which side of that line they sit on.

What "mature agent governance" actually means

Governance maturity is not a feeling about how carefully your team works. It is whether a named person can answer four questions about any agent you run, using evidence rather than memory.

Which agents exist. What each one is allowed to do without asking. What it actually did last Tuesday. Who answers for it when the answer is wrong.

If any of those four takes more than a day to answer, you are in the 79%. That is not an insult. It is the ordinary state of a fast-moving software organisation in 2026.

The four capabilities Deloitte found missing

Deloitte's own write-up names the gaps specifically. Clear decision boundaries separating autonomous action from human-approved action. Real-time monitoring for behaviour that drifts. Audit trails that connect an agent action to an accountable party.

Those are not policy items. They are engineering items with a policy wrapper. That distinction is why so many organisations have an AI policy and still score low.

The risks leaders name are consistent with that reading. Data privacy and security tops the list at 73%, followed by legal and regulatory compliance at 50%, then governance capabilities and oversight at 46%. Every one of those is a control problem, not a model problem.

A policy document is not a control

Here is the position I will defend for the rest of this piece. A written AI policy with no enforcement point scores zero, and it should, because it changes nothing an agent can do at 3am.

A control is something that stops the action. A spend cap enforced in the gateway is a control. A sentence in a Confluence page saying agents should be used responsibly is a preference.

Most self-assessed maturity is measuring the second thing and reporting it as the first. That is the single largest source of error in every number on this page.

The five levels of agent governance maturity

Maturity models are usually vague enough to flatter everyone. This one is built so that each level has an artefact you either have or do not have, which makes it hard to score generously.

Use the middle column to place yourself. Use the right-hand column to work out what breaks if you stay there.

Levels 0 to 2 are the improvising band

Level 2 is where the bulk of competent engineering organisations sit, and it is the most deceptive place on the ladder. Every agent a senior person cared about is governed properly. The problem is the ones nobody championed.

A support agent built by one team in a fortnight carries the same write access to your CRM as the flagship one did. It just never went through the review the flagship got, because the review was a person, not a process.

This is also why agent pilots fail in patterns rather than at random. The failure mode is structural, and it repeats across companies that have nothing else in common.

Levels 3 and 4 are where the 21% sit

Level 3 is the target. It is the point at which governance stops depending on which individual reviewed which project, and starts depending on artefacts that survive that individual leaving.

Level 4 is not the goal for most companies, and I would push back on any consultant who says otherwise. Continuous assurance is worth its cost in regulated sectors and in businesses where an agent touches money or medical records. Everywhere else it buys audit comfort at engineering expense.

My reading of the Deloitte number is that it captures Level 3 and above. That matters, because it means the 79% is not a group of reckless companies. It is mostly Level 2, doing careful work without a common standard.

Score your organisation against 12 questions

Answer each one yes or no. A yes requires evidence you could show someone this week, not an intention. Score 1 point per yes.

Do this with the person who runs the agents in the room, not in a leadership offsite. The gap between what leadership believes and what the runbook says is usually three points.

Scoring. 0 to 3 puts you at Level 0 or 1. 4 to 6 is Level 2. 7 to 9 is Level 3. 10 to 12 is Level 4.

One override, and it is deliberate. If you answered no to question 1, cap your total at 3 regardless of the rest. You cannot govern a population you cannot enumerate, and a strong score on questions 4 through 12 applied to an unknown denominator is a measurement of the agents you happen to know about.

Questions 4 and 12 are the two most commonly skipped, and they are the same question asked internally and externally. The contract version is covered in detail in the breakdown of agent authority clauses, and the failure mode when it is absent is set out in the piece on liability caps and remedies.

Why the other 79% are improvising

Improvising is the right word and it is not a moral one. These teams are making sensible decisions per agent, in the absence of a standard that would make those decisions comparable.

Two specific things break first, and they break in a fixed order.

Most teams cannot count their own agents

The Cloud Security Alliance surveyed practitioners on exactly this, in work commissioned by the identity vendor Strata. Only 21% maintain a real-time registry of their agents. Another 32% keep records that are not real time, 32% plan to build a registry within a year, and 8% have none at all.

Treat that as directional. It is a vendor-sponsored survey and the report does not publish a sample size in the summary, which is a real weakness in the figure and I would rather say so than round it into the argument.

The direction is corroborated by the shape of the problem. Agents get created by the same self-service tooling that created the software sprawl problem before them, which is the same dynamic described in the analysis of what shadow AI actually costs.

Traceability breaks before oversight does

The same survey found only 28% can reliably trace agent actions to a human or system across all environments, while 46% can do it in some environments only. That partial answer is the honest majority position.

Notice the sequence. Oversight appetite is high: 68% rate human-in-the-loop review as essential or very important, and around two thirds require human approval for sensitive data access, system changes and financial transactions.

So the intent is present and the evidence layer is missing. You can approve an action in the moment and still be unable to reconstruct it three weeks later, which is the condition most incident reviews discover on day one. The architecture that avoids it is laid out in the piece on designing human-in-the-loop properly, and the monitoring half in the guide to detecting agent drift in production.

The deadline that would have forced this just moved

Plenty of governance programmes were scheduled backwards from one date. 2 August 2026 was when the EU AI Act's obligations for high-risk systems were due to apply, and it functioned as a budget argument inside a lot of companies.

That date moved. Under the Digital Omnibus agreement, standalone high-risk systems listed in Annex III now face 2 December 2027, and high-risk AI embedded in regulated products moves to 2 August 2028.

The transparency obligations under Article 50 were not deferred. Telling people they are talking to a machine and labelling generated content still applies on the original schedule, with a short grace period for watermarking on systems already in market. The practical checklist for that is in the EU AI Act transparency piece.

Here is the part I disagree with the prevailing take on. Most coverage read the deferral as relief. For governance maturity it is the opposite, because the external deadline was the only thing forcing a Level 2 organisation to fund the step to Level 3.

Remove it and the work competes with roadmap items on internal merit alone. That is a harder argument to win, and it has to be won 16 months earlier than the new date suggests, because the register and the audit trail take quarters to build.

Where this argument is weakest

Three problems with everything above, stated before someone else states them.

Self-reported maturity measures confidence, not control

Every number on this page is somebody's opinion of their own organisation. Nobody audited the 21%. In my experience self-assessment error runs in both directions, and the two errors do not cancel.

Engineering-led companies underrate themselves because they know where the gaps are. Compliance-led companies overrate themselves because they have documents. The 12 questions above are written to require evidence precisely to narrow that spread, and they will not eliminate it.

Two different 21% figures are not one finding

Deloitte reports 21% with mature agent governance. The Cloud Security Alliance reports 21% with a real-time agent registry. Those numbers are close enough to feel like corroboration and they are measuring different things, on different samples, using different definitions.

I have kept them in separate sections for that reason and I would ask you not to combine them. A striking coincidence between two surveys is the most common way a false statistic enters circulation, and the correct response is to name it rather than build on it.

The wider point applies to the maturity model itself. It is a framework I find useful, not a measured instrument, and it has not been validated against outcomes at any company.

Three artefacts that move you from Level 2 to Level 3

If you score 4 to 6 and want to move, the work is smaller than it looks, because three artefacts carry most of the weight. None of them is a purchase.

Build them in that order. The register is what makes the other two scopeable, and a partial register beats a perfect one that arrives next quarter. Identity work sits underneath all three and is covered separately in the piece on non-human access.

Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Note that the third reason is the one this article is about, and that it usually shows up disguised as the second.

A project cancelled for unclear business value is often a project that could never prove its value, because nothing was measured before launch and nothing was traceable afterwards. The pre-launch version of that discipline is set out in the production readiness tests for AI proofs of concept.

Frequently asked questions

Where to start this quarter

Two moves, and the first one takes an afternoon.

Score the 12 questions with your platform lead in the room. Write the number down with the date next to it. A score with a date on it is a baseline, and a baseline is the thing almost nobody has when they later try to prove governance improved.

Then open the register, even as a spreadsheet, and fill in one row per agent you can name from memory. The rows you cannot complete are your actual work plan, and the agents that appear in production but not on your list are the finding you needed.

If you sell software into companies running agents, the same 12 questions are a sales asset. Buyers scoring themselves at Level 2 have a budget problem they cannot yet articulate, and a vendor who can name the artefact they are missing is doing something more useful than describing a feature. The hiring side of that gap is covered in the piece on hiring an agent ops engineer.

Related reading