Who burned all the tokens? Stop surprise spikes in your AI bill

TT is for tokenization. AI models read and write text in small pieces called tokens, and providers bill by the token, for what the model reads and for what it writes. The conversations we are having with prospects and customers alike include discussions of being surprised by AI bills and the lack of awareness about token controls.

This sentiment tends to show up in a finance review or during venting sessions about AI spend, a few weeks after a team connected its agents to a dozen new tools and moved on to the next project. The delegating user innocently prompted agents to do tasks. Some agents looped when they shouldn't have, and the person who delegated the task had no idea it would burn enough tokens to make a bonfire.

Where AI budgets go to die

On each turn, an agent sends the model much more than the task at hand: its instructions, the conversation so far and a description of each tool it could call.

If an MCP (Model Context Protocol) server has 100 tools, the agent sends all 100 descriptions every turn, even when it uses one or none. You pay for every one of those tokens, on every turn, so the bill grows with the number of tools you connect rather than with the work the agent does.

Now add more agents and the people using them. Each model has its own price per token. The invoice only has one total bill amount.

How to see which agent and which request drove your AI spend

To set a sensible limit, you need to know what each team and each person actually uses. Provider dashboards group usage by model and total, so that detail is missing. Agent Trail records who started each agent and what it did, which gives you the numbers to size those limits.

How spend connects to access decisions

When people talk about AI tokens, they may mean two different things. Model tokens are the units an AI provider bills for, and they drive your spend. Access tokens are issued once an agent has signed in, and a service uses them to see who the agent is and who it acts for. Whether that agent can make a particular call is decided separately, with real-time authorization on each request. An agent's loop alternates between the two: a model call decides what to do next, a tool call does it, and the result feeds the next model call.

When a cost spike hits, you need to know who caused it. If every model call and every tool call carries the identity of the agent that made it and the employee it acts for, the spike traces straight to that person, that team and those tool calls. Then you can decide who gets which limits.

See who drives AI spend and limit it with Agent Authority

Agent Authority has visibility into both channels an agent spends through: the requests it sends to model providers and the tool calls it makes. Controlling cost takes both in place. LLM Authority covers the model side and attributes spend to the person, team and agent behind it. MCP Authority covers the tool side, where each tool call adds another model turn to the bill.

• Spend by team, person and model, plus seat utilization (LLM Authority). Usage from coding agents such as Claude Code, Codex, GitHub Copilot and Gemini CLI is attributed to the person and team behind it, so "who spent it" gets an answer with a name attached.
• Budgets and spend alerts (LLM Authority). Set model budgets by team or by person, and get an alert when someone starts using a new model or spend jumps overnight.
• Rate limits on tool calls (MCP Authority). Set a call budget per user, agent instance, MCP server or tool, so a looping agent hits a limit before it runs up the bill. Log-only mode shows what a limit would have stopped before it stops anything, which makes the first week of enforcement much quieter.
• One resource per agent. The connect dialog can give an agent a single resource, such as HubSpot, so it loads fewer tools. That also shrinks the tool tax.
• Webhooks, now available to all orgs. Audit events reach Slack, Microsoft Teams, PagerDuty or your SIEM as they happen, so a refused or throttled agent reaches the team channel before it reaches the quarterly review.

Beyond cost, Agent Authority makes a real-time authorization decision on each tool call, for each agent and the person it acts for, so the same record that explains the bill also shows what each agent was allowed to do.

Helpful questions to ponder before your next AI bill

1. Which agents made the most model calls last month, and who owns them?
2. Is your most expensive model making small yes-or-no decisions that a cheaper one could handle?
3. What stops an agent that loops overnight?
4. Can you test a limit in log-only mode before you turn it on?

Best of luck burning fewer tokens.

FAQ