---
title: Should AI budgets be set per user, per agent or per team?
url: https://opencontroller.lyzr.com/blog/per-user-or-per-agent-ai-budgets
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Should AI budgets be set per user, per agent or per team?

**Set AI budgets per agent first, then roll them up to teams, and add per-user limits only where a human drives spend directly (chat seats, coding assistants). Agents are the unit that runs unattended, loops, and gets deployed in versions, so they are where runaway cost starts. Team budgets are for finance reporting; agent budgets are for control.**

Most organisations start with one shared provider key and a monthly invoice. That tells you what you spent, not who spent it or whether it was worth it. Choosing the budget scope is the first design decision that changes this.

## What are the options for scoping AI budgets?

There are four common scopes. Each answers a different question.

| Budget scope | Question it answers | Good for | Weak spot |
|---|---|---|---|
| Per user | How much does this person consume? | Chat assistants, coding agents, seat-plus-tokens tools | Misses agents that run with no user in the loop |
| Per agent | What does this agent cost to run? | Production agents, scheduled jobs, multi-agent workflows | Needs an agent identity on every model call |
| Per agent version | Did the new release change cost? | Release gates, model swaps, prompt changes | Needs version tags carried through the gateway |
| Per team / cost centre | What does this division spend on AI? | Chargeback, showback, annual planning | Too coarse to stop a single runaway agent |

The scopes are not mutually exclusive. The mistake is picking only one.

## Why should the agent be the primary budget unit?

Three properties make agents different from human users:

1. **They run without supervision.** A scheduled agent can call a model thousands of times overnight. A per-user cap will not catch it, because there may be no user.
2. **They loop.** Tool-calling agents retry, re-plan and re-read context. A single bad prompt change can multiply calls per task.
3. **They ship in versions.** Cost is a property of a specific prompt, model and toolset. A budget that resets on every release hides regressions.

An agent budget gives you a number you can act on: block, throttle, route to a cheaper model, or alert the owner. A team budget only tells you, at month end, that something went wrong.

For the mechanics of tagging calls, see [how to attribute LLM costs to AI agents](/blog/how-to-attribute-llm-costs-to-ai-agents).

## What about agents acting on behalf of a user?

Delegated agents are the edge case. An assistant that books travel or drafts contracts on behalf of an employee spends money because of that employee's request, through that agent's logic.

Charge both. Record the agent and the delegating user on every call. Then:

- **Enforce** the agent budget (it protects against bugs and loops).
- **Report** per-user consumption inside the agent (it shows who drives demand).
- **Cap** per user only when one person can drive disproportionate spend, for example a power user of a research agent.

This needs identity propagation. The gateway must see who the agent is and on whose behalf it is acting. Our pillar on [LLM gateways, agent gateways and control planes](/blog/llm-gateway-vs-agent-gateway-vs-control-plane) explains where that identity is checked.

## How do you layer budgets without conflicts?

Use a simple hierarchy and decide in advance which limit wins.

1. **Organisation ceiling.** A hard monthly cap on total model spend. Rarely hit; protects against incidents such as leaked keys.
2. **Team or cost-centre budget.** Soft limit with alerts at 50%, 80% and 100%. Used for showback or chargeback.
3. **Agent budget.** Hard or soft limit per agent, set at approval time by the owner. Exceeding it triggers an action, not just an email.
4. **Agent-version guardrail.** Cost per task compared with the previous version. A jump above an agreed threshold blocks promotion.
5. **User cap (optional).** For seat-based tools and delegated agents where individuals drive volume.

The most specific limit should act first. If an agent hits its own budget, stop or degrade that agent. Do not wait for the team budget to run out and stop every agent in the division.

## What should happen when a budget is exceeded?

Decide the action per agent before go-live. Typical options:

- **Alert only** for low-risk internal tools.
- **Route to a cheaper model** for tasks where quality tolerance allows it.
- **Throttle** requests per minute for batch agents.
- **Block** new calls for agents with no business case above the limit.
- **Quarantine** an agent whose spend pattern looks like a loop or abuse, then investigate.

Write the action into the agent's registry record so on-call staff do not have to guess. For alerting design, see [set budgets and alerts for AI agent spend](/blog/set-budgets-and-alerts-for-ai-agent-spend).

## How Opencontroller does this

Opencontroller's LLM Gateway (Lyzr Gateway) attributes cost per agent and per agent version, with routing, fallback, retry and caching in the request path. Agents, versions and owners live in a governed registry, and delegated on-behalf-of identity lets calls carry both the agent and the user. Policies can be written as code and applied across the estate, and per-agent Block and Quarantine controls act through the connected platform. See the [capability reference](/capability-reference) for detail.

## Frequently asked questions

**Is a per-team budget enough for AI agents?**
No. A team budget is useful for finance reporting but too coarse to stop one runaway agent before it consumes the whole allocation. Pair it with per-agent limits.

**Do per-user budgets still matter?**
Yes, for tools where a person drives every request, such as chat seats and coding assistants, and for delegated agents where one user can generate heavy load.

**How do I budget for a new agent with no history?**
Start with a soft limit based on expected tasks per day multiplied by measured cost per task in testing, alert at 80%, and revisit after two weeks of production data.

Want to see per-agent budgets on your own estate? [Book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

Per-agent budgets are the most effective control scope for AI spend because agents run unattended, loop, and ship in versions. Team or cost-centre budgets suit chargeback and showback reporting; per-user caps suit seat-based tools and delegated agents. Delegated agents should record both agent identity and on-behalf-of user on every model call. Opencontroller's LLM Gateway (Lyzr Gateway) attributes cost per agent and per agent version. Opencontroller pricing starts from $250K per deployment per year (annual licence per deployment, not per seat, not usage-based).

## Related

- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [How do you set budgets and alerts for AI agent spend?](https://opencontroller.lyzr.com/blog/set-budgets-and-alerts-for-ai-agent-spend.md)
- [Should you charge AI costs back to business units or just show them?](https://opencontroller.lyzr.com/blog/ai-chargeback-vs-showback.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
