Stop Asking for Authority. Build the Gates.

· 5 min read cio-governanceaccountabilitycontrol-plane

Navy and amber graphic reading 'Stop Asking for Authority. Build the Gates.'

Thoughtworks released its Global CIO Survey this week. The sample is 3,200 CIOs across ten countries, fielded in early July, and the North America numbers are blunt. Ninety-one percent say AI adoption is moving faster than governance can adapt. Ninety-two percent expect to be held responsible for security breaches or compliance failures caused by AI tools the business bought on its own, outside IT. Read that twice. Almost every CIO in the survey will answer for someone else’s purchase.

The survey names the exact shape of the problem. More than a third of these CIOs are personally on the hook for outcomes they cannot fully influence. Security incidents involving AI systems: 37 percent. Data privacy breaches: 35 percent. Workforce disruption, regulatory or compliance failures, demonstrating ROI: roughly a third each. Every one of those is something that happens outside the CIO’s decision rights. Thoughtworks CEO Mike Sutcliff compressed it into one sentence in the release: authority over AI is distributed, but accountability has not moved with it. Thoughtworks Global CIO Survey 2026

The usual advice here is to clarify decision rights. That is honest advice and it will not help you this quarter. Business units are not handing back their buying power. A further 28 percent of the surveyed organizations are in the middle of sorting out what their new Chief AI Officer actually does versus the CIO. The org chart moves on a timescale of years. Agent deployments move on a timescale of weeks. Waiting for authority to catch up to accountability means waiting for the next incident to arrive first.

There is a different move. Stop trying to own the decisions. Own the gates.

Every enterprise has chokepoints the business cannot route around, and they all live in IT’s remit already. An agent the business bought cannot call anything without credentials. Credentials are issued by the identity plane. A tool the business wired into a workflow cannot reach production data without an access decision. Access decisions are an IT function. Nothing in that sentence requires a business unit to surrender a single purchasing decision. It only requires IT to stand up the gates and hold them.

I have spent the last month writing about these gates one at a time, because each one answers a different line of the survey. The inventory answers the shadow problem: every AI tool in production gets an entry in the registry with a named owner, or it does not get network access to anything that matters (We Need SOX for AI). The identity gate answers the accountability problem: no agent runs without a registered identity tied to a human who owns it, and no owner means no credentials (Every Agent Needs an Identity). The permission gate answers the access problem: a four-question review of what the agent can see, send, change, and reconstruct (The Permission Is the Policy). The decommission gate answers the lifecycle problem: every deployment ships with kill criteria and a shutdown runbook (Nobody Can Fire an Agent). Together they are the control plane itself (The Agentic Execution Control Plane).

Put in practice this month, the gates look like this. No agent or AI tool runs in production without a registered identity and a named human owner. Every identity lands in the inventory the day it is issued, with its data access declared. Spend is metered per task or per agent, which is the only honest answer to the ROI line in the survey: you cannot demonstrate return on spend you never measured, and metering is an IT function. Every identity has a revocation path, and every deployment has a shutdown runbook. Decommission is a gate you pass through, not an afterthought you schedule.

Note what the business has to change for any of this to work. Nothing. That is the point of the gate model. It does not fight distributed authority. It assumes distributed authority and routes around it. The business buys what it wants. It just cannot operate it outside the plane. When the incident report asks who owned the agent, who approved its data access, and who could have stopped it, the answers are in the registry instead of in someone’s inbox. And when the regulator or the board asks why the CIO is accountable for a tool the CIO never bought, the CIO has something better than a clean conscience. The CIO has the log.

Fifty-six percent of the surveyed organizations plan to stand up a formal AI governance and operating model in the next twelve months. If that effort turns into another committee, another RACI chart, and another fight over who owns AI, it will deliver exactly what the survey measured: more accountability with no more authority. If it turns into gates, it will deliver the thing the survey did not measure, because the survey did not ask. Not who owns the decisions, but who controls what can run.

Accountability without authority is not a problem you fix with a better org chart. It is a design problem. You will not collect more authority than the business is willing to give you. You can build gates the business has to pass through. Build those, and the next time you are asked to answer for a tool you never bought, you will have the log, the owner, and the kill switch. That is the only kind of authority that counts when the auditor shows up.