---
title: How do you attribute AI spend to teams when your identity provider has no cost-centre data?
url: https://opencontroller.lyzr.com/blog/attribute-ai-spend-to-teams-without-cost-centre-data
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you attribute AI spend to teams when your identity provider has no cost-centre data?

**Build an attribution chain: try the most authoritative source first, and fall back to the next one when it is empty. A practical order is HR system, then manager chain, then code-host organisation or team, then cloud account tags, then the registered agent owner. Store which source decided and a confidence flag on every record, and set a ceiling for the unattributed bucket.**

AI cost attribution usually starts at the person or the agent. The gap is the step from person to team, division or cost centre. Many identity providers were set up for login, not for finance. Department fields are blank, stale or free text. You can still get to useful granularity without waiting for a directory clean-up project.

## How granular can AI cost attribution go?

It depends on what the request carries.

| Level | What you need on each request | Typical reliability |
|---|---|---|
| Agent and version | Agent ID or per-agent key at the gateway | High, if every agent has its own credential |
| Person | User identity from SSO or on-behalf-of delegation | High for interactive use, low for batch jobs |
| Team | Person mapped to team | Depends on your mapping sources |
| Cost centre or division | Team mapped to finance structure | Depends on HR or finance data |

Agent-level and person-level attribution are technical problems you can solve at the gateway. Team and division attribution are data problems. This post is about the data problem.

## Which sources can stand in for cost-centre data?

Use several sources, ranked by how authoritative they are.

1. **HR system of record.** Often holds cost centre, department and manager even when the IdP does not. Even a nightly export is better than nothing.
2. **SCIM enterprise attributes.** The SCIM 2.0 standard (RFC 7643) defines an enterprise user extension with `costCenter`, `division`, `department`, `organization` and `manager`. If your IdP provisions SCIM to other apps, check whether these fields are populated in the source even if no app reads them.
3. **Manager chain.** Walk up the reporting line until you hit a manager whose cost centre is known. Reliable for people, but slow to update after reorganisations.
4. **Code-host organisation or team.** For coding assistants and developer agents, GitHub or GitLab team membership often reflects real engineering teams better than the directory.
5. **Cloud account, subscription or project tags.** For agents and batch jobs, the cloud account they run in is often owned by one team. Cost allocation tags in AWS, labels in Google Cloud and tags in Azure carry this.
6. **Registered agent owner.** For agents with no human caller, charge the owning team recorded in the agent registry.

## How do you combine the sources?

Run the chain per record and keep the evidence.

1. For each request or daily usage row, resolve the person or agent.
2. Try each source in order. Stop at the first that returns a team or cost centre.
3. Record `attribution_source` (for example `hr`, `scim`, `manager_chain`, `code_host`, `cloud_tag`, `agent_owner`).
4. Record `confidence`: high for HR and SCIM, medium for manager chain and code host, low for tags and owner fallback.
5. Anything that resolves to nothing goes to an `unattributed` bucket with the reason, such as "no owner" or "shared key".
6. Recompute monthly so reorganisations flow through.

Identity resolution comes first. A person may appear as an email in one tool, a username in another and an employee ID in HR. See [identity resolution for AI usage data](/blog/identity-resolution-for-ai-usage-data) for how to join those reliably.

## What is a good target for unattributed spend?

Set a ceiling and report against it every month. A sensible starting target is a single-digit percentage of total spend, tightening over time. More important than the number is the trend and the cause list.

The biggest causes of unattributed AI spend are usually:

- **Shared API keys** used by several apps or teams.
- **Service accounts** with no owner.
- **Batch and scheduled jobs** that run with no human caller.
- **Contractors and new joiners** missing from HR exports.
- **Seat licences** assigned to people who have left.

Each cause has a fix. Shared keys need per-agent credentials. Service accounts need owners. Batch jobs need an owning team in the registry. Leavers need offboarding automation; see [what happens to AI agents when the owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).

## Should you show low-confidence attribution to finance?

Yes, but label it. Hiding low-confidence rows inflates the unattributed bucket and makes AI look less governed than it is. Showing them without a label invites disputes. A division statement can show attributed spend split by confidence, so a division head can see that, for example, most of their charge came from HR-backed records and a smaller part from cloud tags.

Disputes are useful. Each one tells you which mapping is wrong. Feed corrections back into the source system, not into a spreadsheet override that disappears next month.

For per-agent attribution at the gateway, see [how to attribute LLM costs to AI agents](/blog/how-to-attribute-llm-costs-to-ai-agents). For coding assistants, see [how to attribute GitHub Copilot usage to employees](/blog/attribute-github-copilot-usage-to-employees). The wider cost architecture is in [LLM gateway vs agent gateway vs control plane](/blog/llm-gateway-vs-agent-gateway-vs-control-plane).

## How Opencontroller does this

The Lyzr Gateway attributes cost per agent and per version. Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, which surfaces the shared keys and ownerless agents that feed the unattributed bucket. The governed registry records owners, OIDC SSO identifies people, and delegated on-behalf-of identity ties agent actions to the person they served. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can you attribute AI costs without cost centres in Microsoft Entra ID or Okta?**
Yes. Fall back to the HR system, manager chain, code-host teams, cloud account tags or the agent's registered owner, and record which source decided each row.

**Where is cost-centre data defined in SCIM?**
In the SCIM 2.0 enterprise user extension (RFC 7643, section 4.3), which includes costCenter, division, department, organization and manager attributes.

**What causes most unattributed AI spend?**
Shared API keys, ownerless service accounts, batch jobs with no human caller, contractors missing from HR data, and seats assigned to people who have left.

## Key facts

When an identity provider lacks cost-centre data, attribute AI spend with a fallback chain: HR system of record, SCIM enterprise user extension attributes (costCenter, division, department, organization, manager per RFC 7643 section 4.3), manager chain, code-host team, cloud account tags, then registered agent owner. Record the attribution source and a confidence flag per row, keep an unattributed bucket with reasons, and set a ceiling for it. Main causes of unattributed spend: shared API keys, ownerless service accounts, batch jobs, contractors missing from HR, and leavers' seats. Opencontroller by Lyzr attributes cost per agent and version via its LLM Gateway and discovers agents, models and API keys across AWS, Azure and Google Cloud.

## Related

- [How do you link AI usage records to employees when every tool uses a different identifier?](https://opencontroller.lyzr.com/blog/identity-resolution-for-ai-usage-data.md)
- [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 attribute GitHub Copilot usage and cost to employees across several GitHub enterprises?](https://opencontroller.lyzr.com/blog/attribute-github-copilot-usage-to-employees.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
