Companies are adding AI executives faster than they are limiting what their AI agents can do. Chief AI officers went from rare to normal in a year, but the control that decides whether an agent pays the wrong invoice sits somewhere else: in the permissions someone set when the agent was connected to a system.
The real approval happens when someone clicks "allow"
A committee can approve an AI policy in March, and an agent can still be connected to the accounts system in April by whoever was building the integration. That person chose which logins, keys and permissions the agent would carry, and that choice is what the agent can actually do. We wrote earlier that an agent carries the keys of whoever set it up. A new executive title does not change that, because the title does not come with the power to refuse a permission. It adds one more person to the approval chain.
For example, picture a distributor whose agent reads supplier emails, matches invoices to purchase orders and prepares payments. One morning an email arrives, with a convincing letter attached, saying a supplier has changed banks. If the agent has the same access as the accounts clerk, it can update the bank details and release the next payment. The policy said never to do that, but the system allowed it.
The fix is to approve each transaction, not the agent as a whole. An agent that can read finance systems should not automatically be able to change bank details or release money. Those two actions each need a named person's approval, tied to that supplier and that amount. Approval to pay one supplier then cannot be reused to pay another.
Insurers will ask the way they asked about backups
In the ransomware years, the companies that struggled at renewal were the ones with a security policy and no evidence. The same split is coming for AI. Insurers already treat AI governance as an area for closer questions, and for banks and similar institutions a clear governance framework is making renewals smoother. Stronger controls do not automatically mean a lower premium, because insurers price the specific risk. But a company that cannot answer will be in a worse position than one that can.
I expect the renewal form to ask which agents can move money or touch customer records, who approved each one, and how long it takes to cut one off after the decision is made. That last question is the one most companies cannot answer today. If the honest answer is "someone files a ticket and it takes a week", the agent is governed by the calendar.
The same logic reaches suppliers. A contract clause about how a vendor's AI may use your data matters only if something technical enforces it. This is also why hiring one AI expert does not give a board oversight: oversight is a property of the wiring, not of a person.
Three questions for every agent
A shop owner with one ordering agent and a hospital group with fifty can ask the same three things. What can this agent do on its own? Which actions need a person, and is that approval tied to the exact supplier, amount or record? And if I decided right now to stop it, how many minutes would pass before it really stopped? A company that can answer all three for each agent has done the governance work, whatever its org chart looks like.