How to know where your AI agent money goes
Your business has a new monthly expense: AI agents. Engineers run them, and so do designers, project managers and your support team. Soon everybody will. Across the whole organization the bill can reach tens of thousands a month. Maybe you are an agency and want to pass that cost on to your clients. Or you build a product, and your team tells you AI makes them much faster. But do you know what each project or feature actually cost to build? Did anyone forecast it, and was anyone held accountable for it? Which teams and people stayed within their budget? Which AI agents run over the most?
The answer is a method borrowed from how companies already track people's hours. Every agent reports each token it spends and the work it spent it on, like a timesheet. You add it all up at the end of the month and check it against your providers' bills. Every cost lands on a ticket and on the person who answers for the agent, next to what you planned. Then you plan the next month with what you learned from the previous months. Think of it as cost accounting for AI agents. Here is how it works in four steps.
Plan
Set a token budget for each ticket, the way you would estimate hours in an estimate sheet or story points in sprint planning. Or accept a token forecast built from similar past work.
Record
Every agent reports token entries: tokens, cost, model, start and end time, and a summary of the work. Cloud sandbox and CI agents report on their own once they are set up. Each one has a person who answers for its cost. In local sessions on a laptop, people assign a ticket key and confirm the summaries before entries are uploaded. Full transcripts never leave the machine.
Reconcile
Rules attribute each token entry to its ticket, for example by the ticket key in a branch name. At the end of the month both sides must match: the cost attributed to projects and tickets, and the bills your providers actually sent. Then the month is closed, and finance gets a locked cost sheet.
Learn
See actual against token budget for every ticket, person and provider. Overruns and underruns both get a look. The next token forecast is the median cost of similar past tickets, so it moves toward what the work really costs.
I've built software for clients for fifteen years, and I built Token Controller because I needed these answers for my own managers and clients. It gives the people who run agents an easy way to show what the work was, and the people who pay a number they can check. Claude Code, Codex, Gemini CLI and Copilot CLI are supported, and any other agent can report through the API. It is free for one person, 200 USD or 200 EUR a month per organization on the Team plan, and free to self-host under the AGPL.
For the people who run the agents, it takes almost no effort. Cloud and CI agents report on their own, and laptop sessions are connected to tickets through the Claude Code mod.
There is more when you need it: project and organization overhead, passing cost through to clients, a locked cost sheet for finance, and self-hosting on your own servers.
You can start alone in a few minutes at tokencontroller.com and add your team when the first month closes. If you answer for AI spend, I'd like to hear how you handle your AI bill today. Write to me and I'll show you how it works.