How do you stop an AI agent from overspending?

Give the work a budget rather than giving the agent a rate limit. When the budget is exhausted the work stops and raises an escalation to a human, instead of continuing to spend and reporting the overage afterwards.

Isn't a spend cap enough?

A provider spend cap protects your card, not your project. It cannot tell the difference between one initiative burning its allocation and another running normally, and when it trips it takes everything down at once with no record of what was worth continuing.

Who approves a budget increase?

A named human. The escalation reaches a person in Slack, Teams, or Telegram, and their approve or reject is recorded as a decision attributed to them in an append-only ledger.

Budget controls for AI agents.

The failure is rarely one runaway loop. It is a dozen agents each doing something individually reasonable, and a bill at the end of the month that nobody can attribute to a decision.

01 · Why spend caps don't solve it

A provider cap protects your card, not your project

The usual first move is a hard cap at the model provider. It is worth having, and it is not a control.

It has no idea what the money was for. One initiative burning its entire allocation and another running normally look identical.
It fails everything at once. When the cap trips, every agent stops — including the work that was going fine and was worth finishing.
It arrives after the fact. You learn the ceiling existed by hitting it, with no record of what should have been allowed to continue.

02 · Budget the work, not the agent

The unit that should carry a budget is the work package

Rate-limiting an agent constrains the wrong thing. What has a purpose, an owner, and a value is the work — and that is what should carry a ceiling.

Every work package has an allocation

Work is created with a budget attached. Spend is tracked against that allocation as agents report progress, so consumption is visible per initiative while it is happening rather than at invoice time.

Exceeding it raises an escalation, not an overage

When work approaches or exceeds its ceiling, it stops and raises an escalation. That escalation reaches a human in Slack, Teams, or Telegram with the context attached: what was being attempted, what has been spent, what evidence exists so far.

The increase is a decision with a name on it

Approve or reject is recorded as a decision attributed to the person who made it, appended to a ledger that cannot be edited afterwards. Six months later, "why did this cost what it cost" has an answer with a human at the end of it.

The point is not to make spending hard. It is to make continuing to spend a decision somebody makes on purpose, once the original assumption has been proven wrong.

03 · What this is

Budgets are one organ, not the whole body

Budget enforcement on its own degrades quickly into an approval queue nobody reads. It works when it sits alongside the rest: independent verification so spend maps to attested work rather than activity, validation that ratchets so you are not paying twice for the same regression, and a ledger that makes the whole thing reviewable afterwards.

That combination is what Provostry is — an operating system for AI organizations. Budgets are the organ you feel first, because cost is usually the first thing that gets someone's attention.

04 · Questions

Common questions

How do you stop an agent from overspending?: Budget the work rather than rate-limiting the agent. When the allocation is gone, the work stops and asks a human instead of continuing and reporting the overage later.
Isn't a provider spend cap enough?: It protects your card, not your project. It cannot distinguish valuable work from a runaway, and it takes everything down together.
Who approves an increase?: A named human, in the tool they already use, with the decision recorded against them.

05 · Try it

Put a ceiling on your first work package

Read the integration docs →

Related: an audit trail for AI agents.