Skip to main content

How to set token budgets and spending limits for AI agents.

A spending limit an agent cannot exceed sits in the path its calls take. How to set caps by organisation, team and agent, with a named approver.

Topic
AI cost
Product
Novie Control
Format
Guide
Reading time
7 min
Published
2 October 2026
Updated
29 October 2026
In short

To set a spending limit an AI agent cannot exceed, put the cap in the path its model calls take, where it can stop a call before the call is paid for. Set a ceiling for the organisation, an allocation for each team, and a cap for each agent and each person, in dollars per period. For each cap, write down what happens when it is reached and which person can approve an increase, before the first agent runs.

Token budget, spending limit, cap.

A token budget is the number of tokens an agent may use, per task or per period. A spending limit is the same idea in dollars. A cap is the point at which something happens: the call stops, slows, steps down to a cheaper model, or waits for a person.

Set limits in dollars. Tokens are priced differently by every model and every provider, and differently again for input, output and cached tokens. A limit of a million tokens means one amount on a small model and many times that on a frontier one. Dollars are what the budget is written in, so they are what the cap should hold.

TakeawaySet limits in dollars, because tokens are priced differently by every model, and a cap is the point at which something happens.

Why agents overspend.

A person using a chatbot spends in small, visible steps. An agent spends on its own, at machine speed, and nobody is watching the meter at three in the morning. The usual causes are ordinary.

  • Loops. An agent retries a step that keeps failing, and every retry is a paid call.
  • Context that grows. Each turn carries the whole conversation, so the cost per step rises as the task goes on.
  • The default model. A workflow built on the largest model during a demo keeps using it in production for work a smaller model would do.
  • Agents calling agents. One request fans out into many calls across several models, and the total is visible nowhere.

Agents also use far more than a person does for the same job. The FinOps Foundation reported in May 2026 that reasoning and agentic workloads consume five to thirty times more tokens per task than the equivalent chat interaction.

TakeawayAgents spend on their own at machine speed, through loops, growing context, oversized default models and fan-out, and use five to thirty times the tokens a chat does.

Why caps became a board topic this year.

In June 2026, Uber capped spending on AI coding tools at $1,500 a month per employee for each tool, after its chief technology officer said the company had used its full-year AI budget in about four months. Uber confirmed the cap, TechCrunch reported.

At the end of July, the Financial Times reported internal Amazon presentations in which staff said AI had caused budget overruns that sometimes took months to detect.

TechCrunch, June 2026; Financial Times, July 2026.

These are not isolated. In the 2026 State of AI Cost Governance survey of 396 enterprises, fielded in April and May 2026, a third of companies had imposed emergency spending freezes after AI costs surprised them. KPMG's Q2 2026 AI Pulse Survey found 36 percent of large US enterprises had direct token or usage controls in place.

KPMG, AI Pulse Survey, Q2 2026.

TakeawayUber capped coding tools after its year's budget went in four months, Amazon found AI overruns that took months to detect, and 36 percent of large US enterprises have usage controls.

Start with the limits your providers already have.

Most providers now offer a cap of their own. OpenAI lets you set a monthly hard limit for an organisation or a project, Anthropic for an organisation and each workspace, and Claude Enterprise, ChatGPT Enterprise, Cursor and GitHub Copilot for members and groups. Turn them on as a backstop, a little above planned spend. Each stops at one provider's edge, and none sets a limit for a single agent, so the levels below sit above them. Our comparison of AI spending limits covers what each provider lets you set.

TakeawayTurn on each provider's own cap as a backstop, then set the levels below across every provider.

Where to set limits.

  1. The organisation ceiling. The total AI spend the business has approved for the period. It should be impossible to exceed without a decision.
  2. The team allocation. Each team's share of the ceiling, owned by its lead.
  3. The agent and person cap. A limit per agent, and per person for tools that spend on an individual's behalf, so one runaway process cannot spend a team's month.
  4. The task limit. A token budget per run, so a single task that goes wrong stops after a known cost.

TakeawayA ceiling for the organisation, an allocation per team, a cap per agent and per person, and a token limit per task.

What happens at the cap.

Choose the behaviour per agent, by asking what breaks if it stops.

  • Stop. Right for experiments and for agents whose work can wait until tomorrow.
  • Slow. Queue new runs until the period resets. Right for batch work.
  • Step down. Route the next calls to a smaller model. Right where a cheaper answer is better than none.
  • Ask. Pause and send the request to the named approver. Right for agents in a customer or revenue path, where stopping costs more than the overage.

TakeawayChoose stop, slow, step down or ask per agent, by asking what breaks if it stops.

Increases: who approves, and how fast.

Every increase has a named approver, a reason and an expiry. A temporary uplift for a month-end run should end at month end on its own. Approval should take minutes, because a cap that takes a week to lift will be routed around, and the workaround is the next ungoverned agent.

TakeawayEvery increase has a named approver, a reason and an expiry, and approval takes minutes or the cap gets routed around.

How to control AI agent costs: five levers.

A cap stops the overspend. Four more levers lower the cost before the cap is reached. In August 2026, Gartner predicted that AI inference costs per agentic workflow will increase more than fivefold through 2028, and named inference tiering and routing as the way to keep returns ahead of cost.

Gartner, August 2026.
  1. Cap in the path. Put the budget where the agent's calls pass, at organisation, team, agent and task level, so it holds before spend is committed.
  2. Route each step to the smallest model that does it well. Keep the largest models for the steps that need reasoning, and send classification, extraction and short drafts to cheaper ones.
  3. Keep context short. Summarise long conversations, retrieve only the documents a step needs, and cache prompts and answers that repeat.
  4. Stop loops early. Set a task limit and a retry limit, so a step that keeps failing ends after a known cost.
  5. Review cost per task monthly. Compare each agent's cost per task with the value of the task, and retire or narrow the agents that cost more than they return.

TakeawayCap in the path, route each step to the smallest capable model, keep context short, stop loops early and review cost per task monthly.

A one-page policy.

  • The ceiling: the organisation's AI spend for the period, and who set it.
  • The allocation: each team's share, and the lead who owns it.
  • The caps: per agent, per person and per task, in dollars.
  • The default at the cap: stop, slow, step down or ask, set per agent.
  • The approver: who can raise each cap, and the expiry every increase carries.
  • The review: when the limits are reset, and who reviews the overruns.

TakeawayOne page holds the ceiling, the allocations, the caps, the default at the cap, the approver and the review.

Common mistakes with spending controls for AI agents.

Most overruns come from a handful of setup choices that look reasonable on the day they are made.

  • Setting caps only in the provider console. Each console stops at its own account, so an agent that calls two providers has two partial limits and no total.
  • Alerting instead of stopping. An alert at 80 percent arrives after the agent has already spent it, and a loop overnight can run past 100 percent before anyone reads the email.
  • One cap for the whole organisation. A single ceiling lets one busy agent use the budget every other team was counting on.
  • Caps with no exception path. When the only way past a cap is to ask engineering to raise it, people share keys or move work to a personal card.
  • Increases with no expiry. A temporary raise for a month-end run that never comes back down becomes the new limit.
  • Counting tokens and forgetting tools. Search, retrieval and other paid tools an agent calls belong inside the same budget as its model calls.

TakeawayPut the cap where the call happens, set it per agent, give it a named approver and let every increase expire.

How Novie Control does it.

Novie ControlAvailable now

Cost control that holds.

Set caps per workload and route funding requests to the right approver.

$2,620

of $3,000 cap.

Sources

  1. TechCrunch, Uber caps employee AI spending after blowing through budget in four months (2 June 2026)
  2. Financial Times, Amazon finds cases of AI causing runaway spending on tech projects (30 July 2026)
  3. Mavvrik and Benchmarkit, State of AI Cost Governance 2026 Report (2026)
  4. KPMG, AI Quarterly Pulse Survey Q2 2026 (24 June 2026)
  5. FinOps Foundation, Token Economics: The Atomic Unit of AI Value (10 May 2026)
  6. Gartner, AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028 (17 August 2026)

Other articles

Frequently asked questions

Can you set spending limits on OpenAI and Anthropic?

Yes. OpenAI lets you set a monthly limit for an organisation or a project, as an alert or a hard limit. Anthropic caps every usage tier and lets you set lower monthly limits for the organisation and each workspace. Both stop at their own edge, so a budget across providers and per agent has to be set above them.

How do you control AI agent costs?

Put a cap in the path the agent's calls take, so it holds before spend is committed. Then lower the cost under the cap: route each step to the smallest model that does it well, keep context short and cache repeats, limit retries, and review each agent's cost per task against the value of the task every month.

How do we set budgets for AI agents?

Set a ceiling for the organisation, an allocation per team, and a cap per agent and per person, in dollars per period. Put the cap in the path the agent's calls take so it holds before spend is committed, and name the person who approves increases.

How do we stop AI agents overspending?

Give every agent a cap and a task limit, and decide in advance whether it stops, slows, steps down to a smaller model or asks a person at the cap. Route its calls through one point that checks the budget before each call is paid for.

What is a token budget?

The number of tokens an agent may use per task or per period. Because every model prices tokens differently, it is safer to express the limit in dollars and let the token count follow from the model the call uses.

Should spending limits be in tokens or dollars?

Dollars. Budgets are written in dollars, and the same token count costs very different amounts on different models. A token limit per task is still useful as a guard against a single run going wrong.

Build the intelligenceyour organisation compounds.