How to track LLM costs per team and project.

01Why the invoice cannot tell you

A model provider bills by API key, account or project. It reports tokens and dollars per key and per model. It has no field for the team that ran the workflow, the project the work belongs to, or the business outcome it produced.

One key usually serves several teams. Marketing, engineering and operations share a credential because it was quicker to issue one than three. The invoice then shows a total and hides who caused it, and finance reconstructs the split from logs weeks later, when the month has already repeated itself.

The gap is common and measured. In 2026, 98 percent of FinOps practitioners manage AI spend, up from 31 percent two years earlier, according to the FinOps Foundation's State of FinOps 2026. KPMG's Q2 2026 AI Pulse Survey found 26 percent of large US enterprises had full, real-time visibility into what their AI systems cost to run. In New Zealand, Datacom's 2026 State of AI Index found one organisation in five has a dedicated AI budget with clear cost tracking.

Attribution is where it breaks. In the Tokenomics Foundation's State of Tokenomics, published in September 2026 from 472 large companies, visibility and attribution of spend was the second biggest challenge, after proving value. In research by the cost-management firm Harness, reported by CIO Dive in July 2026, one organisation in five could trace an unexpected AI cost spike to its source within hours. More than 40 percent still managed AI costs in spreadsheets.

02What to record for every call

The record is a ledger line, written when the call runs. Ten fields cover it.

• When. The timestamp of the call.
• Team. The team accountable for the spend, taken from the caller's identity rather than typed in by hand.
• Project. The piece of work the spend belongs to, as finance names it.
• Workflow or agent. The process that made the call, so cost can be read per outcome.
• Model and provider. Which model answered, from which provider, in which region.
• Tokens. Input, output and cached tokens, counted separately because they are priced separately.
• Price. The unit prices in force at the moment of the call, so a later price change does not rewrite history.
• Cost. The computed dollar cost of the call.
• Credential. The key or account the call used, which is what the invoice will show.
• Budget. The budget line the call drew down, and who approved that budget.

03Where to capture it

There are three places a call can be counted. The rule for choosing is simple: capture it where the call cannot avoid being seen.

1. At the provider. Issue a separate key or project per team. It is quick and it gives a per-team view on the invoice, but it stops at the team: it cannot see the project or the workflow, and keys get shared again within a quarter.
2. In the code. Wrap the model library so every call writes a ledger line. It is precise, but it depends on every team adopting the wrapper in every service, including the ones built before it existed.
3. In the path. Route every call through one point that adds the tags, records the cost and checks the budget before passing the call on. It is one place to maintain, it covers every team by default, and it is where a cap can stop a call before it is paid for.

Most organisations start at the provider and end in the path. The first gives finance a number this month. The second gives it a ledger it can trust.

04Rolling it out in four weeks

1. Week one: list the spend paths. Every provider account, every key, every card charge for an AI tool and every vendor that has switched on an AI feature inside a product you already license. Ask each team separately; the differences between the lists are the first finding.
2. Week two: put the tags in the path. Route model calls through one point that stamps team, project and workflow on each request. Where that is not possible yet, issue one key per team so the invoice splits at least that far.
3. Week three: set the untagged rule. Decide that any call without tags is charged to a named person, the owner of the key it used. Untagged spend with an owner falls quickly; untagged spend labelled shared grows.
4. Week four: reconcile and publish. Match the tagged ledger to each provider invoice, explain the difference, and send the first per-team and per-project report to finance and to each team lead.

05Shared costs and the untagged remainder

Some cost belongs to everyone: the platform itself, evaluation runs, the agent that serves the whole company. Allocate it on usage where usage can be measured, keep it central where it cannot, and write the rule down before the first report so nobody argues it afterwards.

The untagged remainder is a line on the report with an owner and a trend. It should fall every month. When it rises, something new is calling a model outside the path, which is worth knowing for reasons beyond cost.

06What finance should receive each month

• Spend by team, against each team's budget.
• Spend by project, against the business case that approved it.
• The workflows that cost the most, with cost per run and per outcome.
• The untagged remainder and its owner.
• Every budget increase approved in the month, and who approved it.
• The change from last month, with the reason in one sentence.