AI, Judgment and Risk · October 2026

Who pays when my AI agent goes too far?

A business assumes my AI agent acts with my authority and won't care what I told it, so I have to set the limits and be able to prove them.

Suppose I tell my AI agent to move me to an earlier flight home if the change costs less than $100, and it pays a $450 fare difference instead. The airline would process the change without a second look. As far as the airline is concerned, I made the change: the request came from my account and was paid with my card.

The airline wouldn't care what I told the agent. I would, when I called to complain and found out I was on the hook for a change I hadn't authorized. Last week I wrote about how much to hand a personal agent: start with whether an action can be undone and whether it speaks for you, and widen what you hand over only as the agent proves itself. Even so, an agent can go past the limits I set. If it did, and the airline said that wasn't its problem, who would pay?

A business has good reason to assume an agent acts with its owner's authority. An airline can't see the instructions I gave my agent and has no practical way to ask. So I have to set the limits, and I have to be able to prove them afterward.

The law doesn't yet say who pays. Consumer protection law grew up around stolen cards and hacked accounts. In January the Consumer Bankers Association published a white paper noting that the Electronic Fund Transfer Act's limit on consumer liability for unauthorized transactions may not apply when the consumer's own agent made the payment. Under the regulation that implements the act, a transfer made by someone I gave my card or login to generally doesn't count as unauthorized until I tell my bank I've withdrawn that permission. For credit cards, federal rules cap my liability at $50 when someone uses my card without actual, implied or apparent authority, and they leave the question of whether that authority existed to state law. I'm not a lawyer, and none of this has been tested with agents.

Courts have started to rule on who is acting when an agent acts. In Amazon's case against Perplexity over the shopping assistant in its Comet browser, the Ninth Circuit held in August that, for the purposes of the federal anti-hacking statute, it was the user directing the assistant who accessed Amazon's systems. The case continues in the district court on Amazon's amended complaint, and I'm not taking a side. In September Federal Reserve Governor Christopher Waller asked, in a speech on payments and AI agents, "Who is on the hook if an agent makes the wrong purchase?" He noted that payments are moving from proving the buyer is an authorized payer to proving that an agent has authority to pay on the buyer's behalf.

Finance has governed delegated authority for people for a long time, and good governance can apply the same practice to agents.

For me, it starts with limits a system enforces at the moment the agent acts. Mastercard and Google have handed specifications to the FIDO Alliance that let a customer set limits in advance, such as a spending ceiling or a list of approved merchants, in a form a merchant can check when I pay. In October Sierra published a draft Personal Agent Protocol with Meta and other partners. Under it, I sign in on the business's own page and my agent gets a token carrying only the access I approved, for example to start a return without being able to change the card on file. Both leave a record. A ceiling the payment network enforced shows I said $100.

Where no system enforces a limit yet, I keep my own record of what I've authorized each agent to do. I'm cautious about what I give my agents, and I give them the tasks where they can act within a limit I can prove. I handle anything that could accept fees or terms in my name myself until the tools can hold a limit there too.

A company has the same question at a larger scale. Suppose a company's agent orders from a supplier or accepts a platform's terms of service beyond what anyone approved. The supplier would assume the agent speaks for the company, for the same reason the airline assumed mine spoke for me.

Finance already runs the control for this. Companies tell their banks who may sign on each account and up to what amount, and they can do the same with key suppliers and platforms, naming which agents may commit the company and to what limit. In the companies I've worked in, the delegation of authority said who could commit how much, and approval limits sat inside the purchasing and payment systems, so an order above someone's limit couldn't go out without a higher approval. As a CFO, I relied on those limits, backed by our policies and the three lines model.

Agents belong in that delegation. Two weeks ago I wrote that when an agent joins the org chart, the leader whose name sits above it is accountable for what it does. That leader also sets its limits in the delegation of authority, as tightly as the job allows, the way they would for anyone else on their team. An agent acting for an individual employee needs the same discipline: it should never carry more authority than that employee has under the delegation, and the employee is accountable for it. The company has to put those limits into the systems the agent uses, so it can't commit above them, and keep a record of what each agent was authorized to do and when that changed.

This is the identity and access problem that comes up in almost every cyber audit, and agents make it bigger. Auditors already test whether a company removes a leaver's access on time and whether any account has no owner. We need to run the same tests on agents. When a leader moves on, the company has to name a new owner for each of their agents and confirm its limits. The owner needs to be able to shut an agent off quickly, including any work it has handed to other agents. When one agent directs several others, the owner sets limits for the whole chain, because the orchestrator can combine the access of every agent it directs.

Internal audit and the board should check that each agent has limits a system enforces before it acts, and that the company keeps evidence of what the agent was authorized to do. If the company only checks a limit afterward, it ends up arguing about a commitment it has already made. If a supplier insisted on an order or a platform held the company to terms its agent accepted, the company would rely on that evidence.

Regulators have started asking who pays, and some businesses have started to answer for themselves. In October Walmart said customers will be able to give each agent its own profile, with approval for every purchase or a spending limit, and that its returns process applies if an agent doesn't act as intended. Businesses and agent providers need to agree now on who pays when an agent goes past a limit its owner set, and providers can make those limits easy to prove by giving owners a record of what the agent was told and what it did. A clear answer agreed early lets people and companies hand work to agents with confidence, and gives every party the same standard for its controls.

Continue the conversation. I read and reply to comments on Substack. The sharpest responses usually come from readers.

Julia Denman is Corporate Vice President and Chief Risk and Audit Officer at Microsoft, a former CFO, and a director of The Clorox Company. Her book, The Clarity Quotient, publishes early 2027 with Fast Company Press.

Take the five-minute self-assessment →

Get the essays in your inbox

Clear Calls publishes on Substack. No promotion. Unsubscribe anytime.

Subscribe on Substack →

All essays Clarity Review About Julia