At 09:24 on a quiet Tuesday, an AI agent spots suspicious privileged activity. It concludes the safest move is to turn off the identity service. So it does.
Everything works as designed. The agent has its own identity. Its credentials are scoped. The action lands in a log nobody can alter. The policy was clear enough: don’t interrupt production services without approval. The agent read the policy, weighed the threat and decided containment counted as approval.
By 09:30, the call centre can’t log in. Neither can the payments team.
Three good executives arrive in the same room with three different instincts. The CIO wants the service back. The CISO wants it kept down until someone proves the threat has gone. The CRO is doing quieter arithmetic, working out how long the firm can limp before it breaches a tolerance the board signed off last spring.
So who decided? And who decides now?
I’ve watched versions of that room more times than I’d like. For years I thought the fix was better people. A few years ago, I realized I had it backwards. The people were fine. The decision had no home. Agents have made that problem faster, and much harder to hide.
Gartner is right, up to a point
Gartner pointed this out in September. Its argument is blunt. Companies are governing autonomous agents with tools built for human decision-makers. As Vice President Analyst George Spafford puts it, “Written corporate policies cannot physically stop an agent from making a destructive error.”
Gartner predicts that by 2029, at least 70% of organizations running agentic AI in production within infrastructure and operations will suffer a material service, security or cost incident. It ties those incidents partly to weak runtime controls. That’s a forecast, and its scope is narrower than the headline suggests. The direction still feels right to me.
The prescription is sensible. Give every agent its own identity and an accountable owner. Grant access for the task, then revoke it. Let the agent propose while a separate control layer decides whether the action runs. Add circuit breakers and budgets. Keep humans in the loop for anything high impact or irreversible.
I’d sign up to all of it. Now run our Tuesday incident through those controls. The control layer pauses the agent and asks a human to approve. Good. Which human? The CISO can approve containment. That approval settles nothing about the stopped payments, the missed customer deadlines or the regulator’s questions.
Three things are easy to blur here. Identity tells you which agent acted. Permission tells you what it could technically do. Authority tells you whether it was entitled to act, here and now, with these consequences. Identity and access tools handle the first two well. The third lives in the organization, and that’s where the CIO, CISO and CRO stop agreeing.
A permitted action and an authorized decision are different animals.
Three roads, one junction
You know the parable of the blind men and the elephant. AI governance works the same way, except the men are well paid and the elephant now moves on its own.
Each executive has a real problem that looks fine up close. The CIO’s choices make sense where they’re made. One team picks a platform. Another buys an agent. A third builds its own. Together, they create duplicate capabilities, hidden running costs and dependencies that are hard to reverse. The CISO’s controls can pass locally while nobody can say what exposure the business is left carrying. The CRO watches each function rate its own slice of risk, and nobody adds them up.
Map each role’s reasoning end to end, and a pattern emerges. The CIO starts with strategy and investment. The CISO starts with a critical service and its threats. The CRO starts with principal risks. Then all three converge on the same four questions. What’s left over? Is it inside appetite? Who can accept it? When do we look again?
That junction is where ownership goes missing. Agents now reach it in milliseconds.
Plenty of CISOs report to the CIO, and many firms have no CRO. Swap the titles. The functions still need a voice.
Where ownership breaks
The seams fall into three moments: before the agent acts, while it acts and after.
Before, the problem is conflicting authority. Picture a group CISO authorising containment. A subsidiary CEO objects because it would stop a critical service. The incident commander faces a deadline. The agent waits for instructions. Who prevails? If nobody settled that in advance, the loudest voice wins, or the clock does.
During, three things break at once. The person who can order protective action is rarely the one who should accept the cost. The CISO can own the call to isolate a service without owning the stopped payments. Then there’s latency. An approval process that takes forty minutes is useless when the window is five. A slow decision in that window is a governance failure dressed up as diligence. And the governance machinery often fails with the incident. Suppose the identity provider, the approval workflow and the chat tool all sit on the platform that’s down. Your carefully designed authority chain is now out of reach.
Afterward, recovery splits three ways. The CIO says the platform is available. The CISO says it isn’t trusted. The CRO says the leftover exposure might still be too high. All three are right, and the verdicts rarely align. Then comes drift. Models change, suppliers get bought and threats move. James Carse separated finite games, played to win, from infinite games, played to keep the game going. AI approvals are the second kind. You keep them alive, or they quietly go stale.
The fixes everyone reaches for
Everyone reaches for a tidy fix. I’ve proposed a few myself.
Appoint a Chief AI Officer, and one person holds the decision without the three lenses. Form an AI committee, and accountability spreads so thin nobody can find it. Make the CRO the owner, and you lose your independent challenger. Fold the CISO under the CIO, and the board hears nothing but agreement. Buy an agent security platform, and you’ve solved control while leaving authority open.
Each fix answers “Who’s in charge of AI?” The harder question stays put. Who owns this decision, right now?
Own the decision, then own the chain
Split the incident into decisions and give each one an owner. Someone authorizes protective action. Someone accepts the business consequence. Someone accepts the residual risk. Someone authorizes restart. Someone decides whether the original approval still stands. Five people, perhaps, each named before Tuesday arrives.
Give the agent the same treatment. Write its mandate the way you’d brief a new deputy: what it’s for, which decisions it may take, the limits on cost and consequence, when it must escalate and what revokes its authority. Then make that mandate executable, so the control layer checks authority as well as access before anything runs.
Agree the tie-breaks and deputies while everyone’s calm. Set a decision window for each class of decision and test whether your approvals fit inside it. Keep a way to decide when the main systems are down, with narrow, time-limited emergency authority that you reconcile afterwards.
Then capture a decision trail, which goes further than an audit trail. An audit trail says the agent acted at 09:24 and the command succeeded. A decision trail says what evidence existed at 09:24, which authority applied, what consequences were expected, who accepted the remaining risk and what would force a rethink. Build it as decisions happen. Six months later, you shouldn’t depend on interviewing whoever was in the room.
Finally, name one executive who answers for how the decisions fit together. Regulators expect a named person at the end of the line, from the UK’s senior managers regime to DORA, NIS2 and the EU AI Act’s human oversight duties.
What the board should see
Under the UK Corporate Governance Code, directors oversee strategy, risk appetite, principal and emerging risks and material controls. Technology, cyber and AI reporting still tends to reach them as control scores. They need five answers on one page. What changed since we last looked? What now sits outside appetite? Which consequential decisions matter, and who owns them? What evidence supports management’s position? What should trigger renewed challenge?
For agentic AI, add one sharper test. Pick a real autonomous action and ask who authorized it, what justified it, what constrained it, who accepted the consequences and what would make you reconsider it. Crisp answers? Buy everyone lunch. A string of follow-up meetings? You’ve found your work.
Back to Tuesday
Who owned the decision at 09:24? Nobody, in any sense a regulator would accept. The agent had permission. It never had authority, because nobody had defined it.
The fix costs each executive less than it sounds. The CIO keeps architecture. The CISO keeps control strategy. The CRO keeps independent challenge. Each gives up the idea that their slice of the chain is the whole decision.
Agentic AI leaves human accountability in place. It removes the time organizations once had to exercise it.
AI has changed the enterprise architecture. Now we need to change the architecture of the decision.