Adoption outran oversight
Something changed in the last year, quietly, inside a lot of organizations. AI stopped only suggesting and started acting. An agent drafts and sends the email. It moves money between accounts. It publishes to the site, updates the record, closes the ticket — on its own, without a person reading every step first.
That shift is worth being clear-eyed about. When AI only produces a draft, a human is still the last line. When an agent acts, the last line moves to whatever oversight you built around it — and for most teams, that oversight was never actually built. Adoption ran ahead of it.
The numbers say the same thing. In McKinsey's 2026 AI Trust research, only about one in three organizations had reached a governance maturity level adequate for the autonomous agents they were already deploying, and roughly 40% were letting agents reach sensitive data with no human in the loop. Deloitte's 2026 State of AI in the Enterprise reads the same way from the other side: agents are broadly in production; formal governance is the part still catching up.
The risk here isn't the agent. It's ungoverned autonomy — an action taken faster than anyone can see it, in exactly the place where a wrong move costs the most.
Governance isn't a binder — it's a short list of things you can check
"Governance" sounds like a committee and a forty-page document. The version that matters day to day isn't that. It's a small set of concrete capabilities you either have or you don't, most of which map to the international AI management standard, ISO/IEC 42001. You don't need to be certified against it to use its logic. You need to be able to answer, honestly, for the agents you are already running.
Here are the seven we would check first.
1. A register. Can you produce a current, complete list of every AI agent you run — right now, not after a week of asking around? If you can't name them, you can't govern them.
2. Per-agent oversight. For each agent, is it defined exactly what it may do on its own and where a human has to sign off? Oversight designed once for "AI in general" isn't oversight; the line has to be drawn per agent.
3. Logging and an audit trail. For any action an agent took, can you see what it did from a log — and undo it if you need to? Reversibility is what turns a bad action into a recoverable one.
4. Evaluation and drift watch. Was the agent tested before it went live, and is it re-checked over time as its model, its data, or the world around it changes? An agent that was safe in March can drift by September without anyone touching it.
5. A kill switch. If an agent starts acting wrong, can you stop it right now — with a mechanism that works even if the agent doesn't cooperate? This is the one to fix before any other. Everything else assumes you can still pull the plug.
6. Accountability. Does every agent have one named owner who is accountable for it and signs off before its scope, model, or data changes? "The team" owning it means no one does.
7. Incident response. Do you have a short, rehearsed sequence for when an agent has already done something wrong? Writing the plan during the incident is how a small problem becomes a public one.
Where most teams are thin
When we walk teams through these, the same two gaps show up first: no register, and no real kill switch. They are related — you can't reliably stop what you haven't named. Neither is expensive to fix. Both are the difference between "we run agents" and "we can stand behind the agents we run."
None of this is about slowing down. A team that can name its agents, halt one, and see what it did moves faster, because it can say yes to more autonomy without flying blind. That is the whole point of a Trusted AI Culture — Frameworks people actually follow, Best Practices embedded across the team, and Internal Champions who keep it intact as things change.
See where you stand in two minutes
Reading a checklist is one thing; scoring yourself against it is another. We built the Agentic Readiness Snapshot for exactly that — seven questions, under two minutes, scored on screen, no sign-up. It walks the same seven checks above and tells you which zone you are in and where your oversight is thinnest, so your next conversation starts with a number instead of a guess.
Take the Agentic Readiness Snapshot → On the /assessment page, switch to the Agentic Readiness Snapshot. Not running agents that act on their own yet? The broader AI Blind Spot Assessment there is the better starting point.
Sources: McKinsey AI Trust research (2026); Deloitte, "State of AI in the Enterprise" (2026); ISO/IEC 42001, the international standard for AI management systems. Informed by ISO/IEC 42001 — the Snapshot points to where oversight is thin; it is not an audit or a certification.