This is the second piece in a two-part series on AI governance and corporate authority. The first1, laid out the blind spot: agents are running in production without a formal approval model. This piece goes one level deeper.
We’ve already established that most boards are flying blind on this: agents are running in production, few went through formal approval, and the incidents keep piling up. That’s a governance blind spot. What it hides is something bigger.
Somewhere between the pilot and the production deployment, something changed that most boards still have not named. The company stopped using AI to help people make decisions. It started using AI to make decisions.
That shift did not happen in a board meeting. It did not go through a governance committee or a legal review. It happened in a sprint, approved by an engineer with a deadline, and it compounded quietly from there.
The result is that a growing number of companies have transferred meaningful decision rights to machines without the formal process that any equivalent transfer of human authority would have required. A new VP with signing authority triggers board approval, legal documentation, and an updated delegation matrix. An AI agent acquiring the same functional authority tends to get a config file and a go-live date.
At its core, this is a story about where corporate authority actually landed, and how quietly it got there.
Corporate governance was built on a premise that most boards have never had to question: the people who hold decision rights can be held responsible for what they do with them. A CFO who approves a bad acquisition answers to a board. A regional director who exceeds her authority answers to someone above her. The chain of accountability is what makes delegation of authority workable at scale.
AI agents break that chain in a specific way. When an agent executes a decision, whether it is issuing a refund, routing a contract, closing a purchase order, or denying a customer request, the accountability does not transfer to the agent. It stays with the company. Ask most companies who signed off on the agent making that call, under what limits, whether those limits still track current policy, and the answer tends to trail off.
Engineering was never going to solve this on its own. That limitation is structural, not a failure of execution.
Guardrails, approval gates, audit logs, escalation paths, engineers can build all of it well. Thomas Squeo, CTO of Thoughtworks for the Americas, and Matt Kamelman, Innovation Choreographer at Thoughtworks, wrote a detailed piece on exactly this side of the problem. Their article Operating System for Enterprise AI walks through the infrastructure that keeps agents reliable once they’re running: context layers, control mechanisms, evaluation loops that catch drift before it compounds. Worth reading if you’re a CTO trying to get this right technically.
But the question of which decisions a machine is authorized to make on behalf of a company is not a technical question. It is the same question boards answer when they approve an acquisition, set a risk appetite, or define the limits of executive authority. The difference is that boards have not been asked to answer it for AI agents yet. In most companies, it has been answered anyway, by default, by the people closest to the deployment.
That pattern is the problem.
A board that has not addressed AI decision rights has a governance lag, not a technology one. The distinction matters because the remedies point in different directions. More technology investment means better tools, faster pilots, stronger engineering teams. Better governance means someone in the room asking questions that engineering cannot answer: which decisions has the company already handed to machines, who signed off, and do those authorizations still reflect what the board actually wants.
This gets worse over time. An agent approved for a narrow task tends not to stay narrow. Systems get extended, use cases get added, and the original approval never gets revisited. In human roles, that kind of drift surfaces in performance reviews, compliance audits, promotion decisions. In agent deployments, it surfaces when something breaks, often at a scale and speed that human decision-making would not have reached.
Naming the problem is not enough. What companies need is the same thing they already have for every other transfer of corporate authority: an explicit Delegation of Authority model, built for machines instead of people.
What a Delegation of Authority Model for AI Agents Looks Like
Nobody has to invent this from scratch. Boards already run on a simple rule for human authority: bigger impact and less room to undo an action means a higher bar for approval. Apply the same rule here. The only new part is naming, explicitly, which tier of the company signs off at each level, before the agent goes live rather than after something breaks.
Customer service agent with authority to issue refunds. An agent that summarizes support tickets and suggests responses to a human representative carries low risk. A manager can approve that use, with the CTO aware. Every output goes through a person before anything happens. The picture changes when that same agent starts issuing refunds automatically up to a certain threshold, with no human review in the loop. In a high-volume e-commerce operation, a miscalibrated agent can push through millions in unwarranted refunds in a matter of hours, before an alert reaches anyone with the standing to stop it. Who approved that capability? Under what conditions? With what spending ceiling and what audit record? Those answers need to exist before the agent goes into production.
Procurement agent with authority to close supplier orders. At this level, approval belongs to a broader group: the CTO, the CFO, and depending on volume and strategic significance, the executive committee. An agent closing a $50,000 order with a strategic supplier is exercising authority that no one at the operational level would have without sign-off. The difference here is that when the authority was handed to the system, no one signed anything.
HR, legal, and financial agents with authority to communicate decisions or execute transactions. An HR agent that relays hiring decisions to candidates, a legal agent that routes contracts for execution, or a financial agent processing transfers within preset limits are each operating in territory where errors carry consequences well beyond the operational layer. Reputation, regulatory exposure, and direct legal liability for the company are all in play. Each context demands an approval chain proportional to what is at stake.
Deploying an agent with execution authority is not a technical decision dressed up in business language. It is a governance decision, the formal transfer of decision rights to a system, with everything that implies for risk, accountability, and legal exposure. Companies that treat it as anything less are not moving faster. They are accumulating undisclosed liability.
Conclusion
From naming which agents exist to building the model above, the governance work is familiar, even if the subject is new. Define which classes of decisions require board-level approval before being delegated to a machine. Establish thresholds analogous to existing authority matrices. Build in periodic review so that authorizations do not silently expand. Make accountability explicit before something goes wrong rather than after.
None of this requires the board to understand how large language models work. It requires the board to apply the same governance instincts it has always used when authority is being transferred from one party to another.
What matters is whether anyone with board-level authority decided what AI is allowed to decide, and whether that decision has a name attached to it. Whether your company is using AI stopped being the interesting question a while ago.
In most companies, the honest answer is no. The agents went live anyway.
Who approves your company’s AI autonomy level?
The number of AI agents operating inside companies doubled in less than a quarter. A survey of 750 senior technology leaders conducted in April 2026 found that 80.9% of organizations have already moved past the pilot stage into real production. Agents are making decisions, executing actions, and touching critical systems every day.