AI agents are already spending, per call and per token, on behalf of their publishers and clients, without having their own account. In an analysis published on Journal du Net, Charles Kremer, founder of Keyban, a company designing the financial infrastructure for AI agents, advocates for the idea of a per-agent account, organized by mission, which caps spending before the mission and reveals the margin after.
- AI agents already spend money per call and per token, yet they have no account of their own.
- Charles Kremer proposes one account per agent, organized by mission, that caps spending in advance.
- Flat-rate sales leave publishers absorbing cost overruns, and selling credits only shifts the risk to clients.
- A mission account sets an objective and a budget, so the agent works alone and reports only deviations.
The author gives the example of a prospecting agent, sold on a flat-rate basis, which enriches records for a client. After a change in its instructions, the agent re-reads each record multiple times, and its cost per record quadruples. Its access keys remain valid and each call to the model is billed on a usage basis, so nothing stops it. According to him, the publisher only discovers this on the monthly invoice.
This situation is not exceptional. According to an IDC survey, 67% of the companies surveyed exceeded their compute and model budget for their agents by more than 10% over twelve months. A company that exceeds its budget can renegotiate it, whereas the publisher selling the agent on a flat-rate basis bears every overrun on its margin. ‘The problem is not the agent, it’s that it spends without its own account,’ writes the author.
Spending that will multiply
Models are already billed per token, a unit of computation they consume, and the services used by an agent, such as web search, contact enrichment, or company verification, are increasingly sold per call. According to Gartner, by 2028, 90% of business-to-business purchases will be made through AI agents, representing over $15 trillion.
For the author, the driver of this trend is less about new payment methods, which allow settling a few cents per request, than about the very cost of purchasing: comparing, negotiating, getting an invoice approved. The agent reduces this cost, and the unit of purchase shrinks down to the actual need. However, every tiny expenditure incurred by a non-human actor must be attributed to the correct client.
Selling credits is transferring the risk
Clients demand a fixed price, whether it’s a subscription or a price per successful task, while the publisher’s cost varies with each mission. According to ICONIQ, the average gross margin for AI products reached 45% in 2025. In July 2025, Anthropic discovered that a user running its coding tool continuously had consumed tens of thousands of dollars on a $200 plan, and then instituted weekly quotas. In June 2025, Cursor replaced the requests included in its subscription with a usage credit priced at model rates, before apologizing and refunding customers surprised by their bill three weeks later.
For the author, selling credits protects the margin without solving the problem: the risk of overrun shifts to the client, who loses visibility into what they will pay for a given result.
Card, key, or account
Three tools currently allow controlling spending, but none are sufficient. The card pays without knowing for which client, the access key opens a model without seeing services or purchases, and the credit shifts the risk. The account, however, combines the holder’s identity, the money entrusted to it, and all its expenditures, including models, services, and purchases.
This account is organized by mission, with an objective, a budget, and a deliverable. The human validates the intention once, the agent then works alone, and only deviations are reported. In the author’s example, a mission of 200 records with a budget of 100 euros, or 50 cents per record, pauses as soon as the cost reaches 2 euros per record, with an alert, before the budget is exhausted.
