---
title: How do you tie every AI agent to a cost centre when it is approved?
url: https://opencontroller.lyzr.com/blog/tie-ai-agents-to-cost-centres-at-approval
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How do you tie every AI agent to a cost centre when it is approved?

**Make the cost centre a required field in the agent approval request, store it on the agent's registry record, and propagate it to every credential, cloud tag and gateway route the agent gets. If the cost centre is captured once at approval and carried automatically, spend attribution becomes a lookup instead of a monthly investigation.** Keep the form short: a handful of fields that each drive a control later.

Teams that skip this end up reverse-engineering spend from API keys and invoice lines. Teams that over-engineer it build a 40-question intake form that makers avoid. The goal is the smallest form that still makes chargeback, budgets and alerts automatic.

## Which fields should the approval request include?

| Field | Example | What it drives later |
|---|---|---|
| Purpose (one sentence) | "Drafts first replies to tier-1 support tickets" | Scope checks, re-approval when purpose changes |
| Business owner | Named person, not a team alias | Budget alerts, access reviews, incident contact |
| Cost centre | Finance code from your ERP | Chargeback or showback line |
| Monthly budget cap | A currency amount | Soft and hard alerts in the gateway |
| Environment | Dev, test, prod | Separate budgets; dev spend not charged to the business |
| Data classes touched | Public, internal, confidential | Risk tier and guardrails |
| Models and tools requested | Model family, MCP servers | Allow-lists and cost estimate |

Seven fields is enough for most agents. Higher-risk agents can trigger a longer review, as in a [registration questionnaire](/blog/ai-agent-registration-questionnaire).

## How does the cost centre travel from form to invoice?

1. **Registry record.** Approval creates the agent's record with the cost centre and owner. This is the source of truth.
2. **Credentials.** Every API key, virtual key or workload identity issued to the agent is linked to the registry ID, never to a person.
3. **Gateway routes.** The LLM gateway stamps each request with agent ID, version and cost centre, so token spend rolls up without guesswork.
4. **Cloud tags.** Cloud resources the agent runs on carry the same agent ID and cost centre tags. On AWS, user-defined tags must be activated as cost allocation tags before they appear in billing reports.
5. **Export.** Spend data goes to finance in a consistent schema. The FinOps Foundation's FOCUS specification is a useful common format for combining cloud and AI costs.

## What goes wrong without it?

- **Shared keys.** One key used by five agents makes per-agent cost impossible. See [should AI agents share a service account](/blog/should-ai-agents-share-a-service-account).
- **Personal ownership.** When the builder leaves, the cost centre is lost with them.
- **Build and run mixed together.** Coding assistant seats and production agent tokens end up on one line. Keep them separate, as in [separating AI build costs from agent runtime costs](/blog/separate-ai-build-costs-from-agent-runtime-costs).
- **Late tagging.** Tags added after launch do not apply to past spend, so the first months stay unattributed.

## What about agents that already exist?

Run a one-off backfill:

1. Export current agents, keys and models from discovery.
2. Match each to an owner using IdP data and repo ownership.
3. Ask owners to confirm purpose and cost centre in a short form.
4. Park anything unclaimed under a central "unallocated AI" line and give owners 30 days before keys are rotated.
5. Report the unallocated share monthly until it is close to zero.

For agents where finance codes are missing entirely, see [attributing AI spend to teams without cost-centre data](/blog/attribute-ai-spend-to-teams-without-cost-centre-data).

## Checklist: cost-centre-ready approval

1. Cost centre and owner are mandatory fields, validated against your finance list.
2. Approval creates a registry record automatically.
3. Credentials are issued per agent and linked to the record.
4. The gateway tags every request with agent, version and cost centre.
5. Cloud resources carry matching tags, activated for billing.
6. Budget caps from the form become gateway alerts.
7. A change of cost centre or purpose triggers re-approval.
8. Monthly report shows spend by cost centre and the unallocated remainder.

## Showback or chargeback?

Start with showback: send each cost centre a report of its agents' spend for a quarter before moving money. It surfaces wrong mappings without disputes. Compare the models in [AI chargeback vs showback](/blog/ai-chargeback-vs-showback).

## How Opencontroller does this

Opencontroller's governed registry holds each agent, its versions, tools and owner. The LLM Gateway (Lyzr Gateway) attributes cost per agent and per version, and auto-discovery finds agents, models and API keys across AWS, Azure and Google Cloud that were never registered. Policy as code can require owner and cost-centre metadata before an agent is promoted. For how gateways and control planes split this work, read [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should the cost centre belong to the builder or the business owner?**
The business owner of the process the agent serves. Builders change; the business process and its budget usually do not.

**Can one agent be split across cost centres?**
Yes, but split by route or tenant in the gateway, not by percentage guesses. Each route or tenant carries its own cost centre.

**When should the cost centre be re-checked?**
Whenever the agent's purpose, owner or business unit changes, and at each periodic access review.

## Key facts

Tie AI agents to cost centres at approval: seven fields (purpose, business owner, cost centre, monthly budget cap, environment, data classes, models and tools). Flow: registry record -> per-agent credentials -> LLM gateway stamps agent ID, version, cost centre -> cloud tags (AWS user-defined tags must be activated as cost allocation tags) -> export, e.g. FinOps FOCUS format. Backfill existing agents with an unallocated line and 30-day owner window. Start with showback before chargeback. Opencontroller (by Lyzr): governed registry, LLM Gateway cost attribution per agent and per version, auto-discovery across AWS/Azure/Google Cloud, policy as code. Annual licence per deployment, starts from $250K per deployment per year.

## 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)
- [Should you charge AI costs back to business units or just show them?](https://opencontroller.lyzr.com/blog/ai-chargeback-vs-showback.md)
- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
