Product

Your AI agents are already spending. Who’s keeping their accounts?

AI agents spend on your customers’ behalf, without an account of their own. Credits shift the risk onto the customer. One account per agent, organised by mission, caps the cost upfront and reveals the margin afterwards.

·Charles Kremer
Op-ed by Charles Kremer, founder of Keyban, published in the Journal du Net: “Your AI agents are already spending. Who’s keeping their accounts?”

This column was originally published in French in the Journal du Net on November 10, 2026.

Per token, per call, day and night: the AI agents you sell are spending on your customers’ behalf. Yet they have no account of their own, and you discover the drift on the monthly bill. Selling credits fixes nothing: it shifts the risk onto the customer. Every agent needs its own account, organised by mission.

A loop at 3 a.m.

It is 3 a.m. A prospecting agent, sold at a flat rate by a software vendor, is enriching records for one of its customers: executives, company registration numbers, buying signals. The day before, someone changed its instructions. Since then, the agent rereads each record several times before moving on to the next, and its cost per record has quadrupled. Nothing stops it: its access keys are valid, and every call to the model is billed by usage. The vendor will find out on the monthly bill.

The scene is nothing unusual. In an IDC survey, 67% of the companies polled exceeded their compute and model budget for agents by more than 10% over twelve months. A company that goes over budget renegotiates it. The vendor selling the agent at a flat rate pays for every point of overrun out of its margin.

My conviction is simple: the problem is not the agent, it is that the agent spends without an account of its own. We have given software the power to spend by the second, and we still monitor it by the month.

Tomorrow, agents will buy by the cent

And this spending is about to multiply. Models are already billed per token, the unit of compute they consume. The services an agent uses, such as web search, contact enrichment and company verification, are increasingly sold per call. Tomorrow, agents will no longer buy only technical calls, but outcomes: a verified company, a prospecting file, a reviewed translation.

Your agents won’t stop at data. According to Gartner, by 2028, 90% of business-to-business purchases will go through AI agents, more than $15 trillion. Today, your agent recommends and your customer pays. Tomorrow, it will settle small, variable purchases itself on that customer’s behalf, on demand rather than by subscription. That is all the more spending to manage, and to re-bill.

People assume the change will come from new payment methods, which make it possible to pay a few cents per request with no subscription and no sign-up with the supplier. They help. But the real driver lies elsewhere: buying is expensive. Comparing, negotiating, onboarding the supplier, getting sign-off, reconciling the invoice. Nobody buys a piece of data for one euro, because the purchase costs more than the data. So we bundle, we overbuy, and we pay for what we don’t use.

The agent makes that cost collapse. Among suppliers approved in advance, it compares, orders and pays in seconds. The unit of purchase shrinks down to the actual need. The consequence is mechanical: thousands of tiny expenses, committed by an actor that is not human. For the vendor, each of these expenses is either a cost against a price that does not move, or a purchase to re-bill to the right customer. Either way, it must know on whose behalf the expense was committed.

Selling credits is selling the risk

That fixed price is what your customers demand: a subscription, or a price per successful task. Your cost, on the other hand, varies with every mission. All the risk is yours, and the margin is thin: a 45% average gross margin for AI products in 2025, according to ICONIQ.

The biggest players learned this the hard way. In July 2025, Anthropic found that one user, running its coding tool non-stop, had consumed tens of thousands of dollars on a $200 plan. Its response: weekly quotas and, beyond them, extra usage billed at API prices. Cursor went further: in June 2025, it replaced the requests included in its subscription with a usage credit at model prices, and three weeks later it apologised and refunded the users whose bills had taken them by surprise. In both cases, the overrun ends up with the customer.

Selling credits has become the default answer. The customer buys a stock of usage, the agent consumes it, and when a loop runs away, it is the customer’s credit that melts. The vendor protects its margin, but it has solved nothing: it has moved the overrun risk to its customer. The customer loses what subscription software used to guarantee: knowing how much they will pay for what outcome. It is a step backwards.

To guarantee a price for an outcome again, the vendor must know, before each mission, the most it can cost.

No card, no key, no credit: an account

The vendor has three tools for this. Credit, as we have just seen, moves the risk. A card pays, but does not know for which customer: capped at €100, it cannot tell whether those €100 serve customer A or customer B. An access key opens a model, but sees neither services nor purchases, and only cuts off once the budget is exhausted. Each controls only one piece of the spending. None says what an agent spent, for whom, and what it earned.

That is the role of an account. An employee who spends for the company has an identity and an expense report. Every agent that spends for a customer must have its own account: an identified holder, the money entrusted to it, and all its spending in one place, models, services and purchases alike, whatever the payment method. A card pays, a key opens: only an account can account for the spending.

The mission: capped upfront, profitable afterwards

This account is not run payment by payment: nobody will approve thousands of them. It is organised by mission, with an objective, a budget envelope and a deliverable. A human approves the intent once. The agent then works on its own, and everything it consumes, models included, is debited from the mission’s envelope. Only the deviations come back up. At the end, the account reconciles the mission against what the customer paid: cost, revenue, margin.

Back to our 3 a.m. agent, this time with an account of its own and a mission: enrich 200 records for customer X, with €100 set aside, models included, or 50 cents per record. When each record starts costing €2, the envelope would only cover 50 records out of 200. Because the account relates every expense to the records delivered, the deviation is spotted within the first few: the mission pauses and an alert goes out. A simple spending cap would have seen nothing: it would have let the €100 run out on 50 records, then cut the agent off without saying why or for which customer. In the morning, the vendor reads one line on the agent’s account: mission paused, cost per record up fourfold, customer X. It fixes the instructions, restarts the agent and, on delivery, reconciles the spending with what customer X pays: it knows its margin on that customer.

The day an agent makes a mistake, the supplier, the bank and the auditor will ask the same question: was this payment authorised? The bank will see a payment, the supplier an order. Only the agent’s account links them to the mission that authorised them.

Every agent its own account

None of this needs a new protocol or lower fees. It takes one decision: treat every agent that spends for a customer as a mandated employee, and open an account of its own. No company lets an intern spend without a budget, without an expense report and without knowing which customer they are working for. Yet that is what we do every night with our agents.

The next time an agent loops at 3 a.m., it should not cost you a month of margin. It should cost you one line on its account, in the morning.

Charles Kremer is the founder of Keyban, a company building the financial infrastructure for AI agents.