Companies that let AI agents work on their own are finding that the bill is hard to predict, and the first fixes are arriving as limits, not reports. In August Google added a monthly spending cap that pauses a project's agents when the limit is reached. The idea behind it is older than AI, and it explains what a sensible company should do next.
Why an agent bill moves
A single chat answer is small and bounded. An agent is a loop: it reads a request, calls a tool, reads the result, decides what to do, and calls again. Each call usually carries the whole conversation so far, so a job that goes wrong and keeps retrying gets more expensive for as long as it runs. The person who started it sees only a task that is still working.
There is a second meter as well. Every prompt, tool call and model response also produces a trace that monitoring tools store and charge for, so the monitoring bill grows with the model bill. A loop that runs all night is paid for twice.
The phone bill answered this first
When the EU dealt with roaming bill shock, it made three choices. The network enforced the limit at the moment of use, so nobody found out from the invoice. A warning came at 80% of the limit, so a traveller could act before being cut off. And the cut-off applied unless the customer chose otherwise, because few people set a limit for themselves before a holiday. Customers could still opt out of the limit if they wanted to.
Google's cap in Gemini Enterprise has the first choice built in: when a project reaches its limit, its agents pause and other projects keep running. We covered that cap in August. A cap you have to set protects only the teams that remember to set it, which is why the phone rule made it the default.
A budget for each workflow
Picture a distributor whose operations team builds an agent to answer supplier emails. A bug makes the agent retry a failed order lookup every few seconds all night. With only a monthly invoice, the finance director learns about it three weeks later, as a line she cannot explain. Now give that workflow a budget of $300 a month, a warning at $240 and a stop at $300. The loop ends on the first night, and a named person on the operations team gets a message instead of a surprise.
A hard stop has a price of its own. An agent that halts halfway through processing orders leaves work half done. So the choice depends on the job. Experiments should stop hard. An agent that serves customers should warn its owner loudly and finish the task it is on. The phone rule allowed the same split, because customers could opt out of the cut-off.
The budget also has to come from the right person. Whoever starts a workload is the only one who knows what it should cost, which is why some companies now give each employee an AI budget. A budget set by finance, after the fact, is a review. A budget set by the person who starts the job is a limit.
The rule has two parts. Before any agent goes live, a named person owns a monthly number for it. And when you buy monitoring for agents, ask the vendor how it prices, samples and retains traces. Keeping a sample of traces costs less than keeping all of them, and that bill grows with every action the agents take.