Who pays for artificial intelligence agents in a company? In an analysis dedicated to the question, its author argues that the usual answer, the IT department, is no longer sufficient. He advocates for usage-based billing, supported by a gateway that tracks each call to the models.
In most large corporations, generative AI tools are managed by the IT department, then billed to subsidiaries and business units based on the number of users. This model, based on the number of licenses, is the most widespread and remains logical for conversational assistants. For agents, the author believes it no longer works.
An assistant is easy to account for
An assistant corresponds to a license assigned to a person. We know who uses it and we bill their unit, which makes the calculation simple. From an accounting standpoint, the model holds.
It does, however, have a blind spot. According to the author, an assistant acts as a multiplier: it enhances already high-performing employees and brings little to those who were not. In a gradual rollout, licenses first go to those who request them, i.e., the most advanced, so the gap between colleagues widens without anyone having decided it. The per-license model does not create this bias, but it does not allow it to be seen.
Agents are not tied to a person
An agent functions differently. It does not consume a subscription but resources on a usage basis, such as tokens (the units of text processed by a model), API calls (the interfaces that allow software to request a model), and human supervision time. It works for a specific function, such as accounts payable, recruitment, or support, and often calls upon other agents.
The author takes the example of a supplier invoice processing agent. It receives a document, calls an extraction agent, then a compliance control agent, before submitting a result to an accountant for validation. Who is the user then? The accountant who validates, the supplier who sent the invoice, or the finance department that commissioned the agent? Billing such a system per person amounts, in his view, to dividing an apartment building’s electricity bill by the number of residents rather than by each apartment’s meter.
The gateway, a tool for traceability
This meter exists in the form of an AI Gateway, a single gateway through which all calls to models and agents pass. Each call is traced there: we know which agent consumed what and we can trace back to the person and the unit originating the request.
This traceability allows, according to the author, to assign the cost of an agent to the function that consumes its work, with a business manager accountable for its performance, rather than to an IT budget line. He adds that many companies are currently building this type of gateway and that the billing rule must be set at this stage, otherwise each unit will defend the model most favorable to it.
The author, however, nuances his point. An agent is not a full-time equivalent, as it absorbs tasks rather than an entire position. Furthermore, according to the Oliver Wyman Forum (The Reality Check, October 2026), one hour of AI-assisted work requires on average 22 minutes of verification, a time that is part of the agent’s real cost without appearing on any invoice. Finally, some departments might prefer a diffuse IT line, which does not expose them to having to defend the performance of each agent.
