How to allocate an Anthropic or OpenAI bill across departments (2026)

Allocate a shared AI bill on the driver that caused each part of it: seat counts for subscriptions, and metered usage by workspace, API key or project for tokens. Use a documented fallback weight, such as engineering headcount, only for the part you can't trace, and re-run the allocation every month from the provider's usage export instead of booking the whole invoice to IT.

That matters because one Anthropic or OpenAI invoice usually covers very different things: production inference behind your product, an internal support agent, engineers' experiments and a few hundred employee seats. Booked to one IT line, the bill tells nobody anything. Split by driver, it tells each department what its AI use costs and gives finance a number it can defend.

Bottom line: Split the invoice into seats and usage first. Allocate seats by each person's department from the HRIS, usage by the workspace, API key or project it ran under, and spread only the untraceable remainder with a fallback weight you write down. Set up one key or workspace per team so next month's allocation runs itself.

How to allocate AI costs in five steps

The method works the same way for Anthropic, OpenAI, Cursor and most other AI vendors:

1. Split the invoice into seats and usage. Seats are fixed per-person subscriptions. Usage is metered: tokens, agent runs or overage charges.
2. Allocate seats by the seat roster. Export the list of people with seats and match each one to a department and cost center from your HRIS.
3. Allocate usage by where it ran. Export usage or cost by workspace, API key or project, and map each one to the team, application or cost center that owns it.
4. Spread the remainder with a documented fallback. Usage that ran under a default workspace or an untagged key gets a fallback weight, such as engineering headcount or each department's share of traced usage.
5. Reconcile to the general ledger. The allocated total must equal the invoice amount booked that month, before you report anything.

Then repeat it monthly with the close. Some exports only go back 90 days, so a monthly habit also keeps the data from expiring before you use it.

Which allocation driver fits each AI cost?

Match the driver to the way the cost is billed. Here's the standard mapping:

Allocation answers which department pays. Classification answers where it sits on the P&L: customer-facing inference is COGS, employee tools are operating expense and experiments are R&D. Do both, in that order; our guide to whether AI spend is COGS or OPEX covers the classification side.

Where does AI spend hide before you allocate it?

In more places than the one invoice you're looking at. Our budgeting for AI spend guide lists six sources to collect before you allocate anything:

• Seats and subscriptions: employee tools and internal copilots, found in AP and software invoices.
• API and model usage: inference and provider charges, found in each provider's billing and usage console.
• Coding tools and development: copilots, experimentation and model work, found in engineering software, cloud and expense data.
• Other AI tools and software: specialist apps, found in AP, procurement and employee cards.
• AI infrastructure: compute, storage and data platforms, found in cloud and warehouse billing.
• Agents and workflows: automated internal or customer work, found in agent platforms, workflow logs and provider usage.

Allocate each source with its own driver. A company that allocates only the Anthropic invoice and ignores the rest will understate every department's AI cost.

What usage data does each AI provider export?

Enough to allocate most of the bill, as long as you set things up to use it. As of November 2026:

• Anthropic (Claude API): usage and cost reports can be grouped by workspace and API key, and pulled programmatically with an admin key. Workspaces are the natural unit for cost attribution, so one workspace per team or application makes the allocation direct.
• Claude Team and Enterprise seats: owners can export a spend report from the Analytics page, with one row per person and model, covering up to 90 days. Match each person to a department.
• OpenAI (API): usage is organized by project, each with its own API keys, and the usage dashboard can be filtered by project. One project per team or application gives you a clean split.
• ChatGPT Enterprise: workspace analytics include user-level exports, which you can match to departments the same way as Claude seats.
• Cursor (Teams): spend is tracked per user in the admin dashboard, including usage beyond each seat's included amount, so coding-tool overage can go to the engineer's team rather than a shared pool.

Our walkthrough on getting spend data out of Claude, ChatGPT and Cursor shows exactly where each export lives and which access level you need.

Worked example: allocating a $60,000 AI bill

Here's an illustrative month for a 150-person software company with four AI charges:

• Anthropic API, $30,000. The production workspace used $18,000, the customer support agent's workspace $5,000 and the engineering sandbox $4,000. Another $3,000 ran under the default workspace with no owner.
• Claude seats, $15,000 for 150 people: Engineering 60, Sales and marketing 40, G&A 30 and Support 20, so $6,000, $4,000, $3,000 and $2,000.
• Cursor, $9,000, per-user spend: $8,400 from engineers and $600 from two analysts on the data team in G&A.
• OpenAI API, $6,000. A sales research project used $2,500 and a finance close project $1,500; $2,000 ran on an old shared key.

The untraceable remainder is $5,000 ($3,000 plus $2,000), about 8% of the bill. The documented fallback spreads it by AI seat count: Engineering 40% ($2,000), Sales and marketing about 27% ($1,333), G&A 20% ($1,000) and Support about 13% ($667).

The result:

• Product (COGS): $18,000
• Engineering (R&D): $20,400 ($4,000 sandbox, $6,000 seats, $8,400 Cursor, $2,000 fallback)
• Sales and marketing: $7,833 ($4,000 seats, $2,500 research project, $1,333 fallback)
• Support: $7,667 ($5,000 agent, $2,000 seats, $667 fallback)
• G&A: $6,100 ($3,000 seats, $600 Cursor, $1,500 close project, $1,000 fallback)

That adds back to $60,000. Next month, the fix is obvious: give the default workspace and the old OpenAI key owners, and the fallback shrinks toward zero.

How should you allocate AI costs that serve customers?

By product line or customer segment, not by department. The production inference in the example above doesn't belong to any team's operating budget; it's a cost of delivering the product, so it goes to COGS and, if you report margins by product, to the product that triggered it. Give each AI feature or product its own workspace or project, and the export splits COGS by product automatically.

This is the allocation investors care about most, because it moves gross margin. Software companies held a median gross margin around 80% in our 2026 SaaS gross margin benchmarks, and isolating AI inference inside COGS is how you'll see compression early if model costs rise faster than revenue.

Should you use showback or chargeback for AI costs?

Start with showback. The FinOps Foundation treats showback, reporting costs to the teams responsible for them, as required in any cost practice, and chargeback, posting those costs to the teams' official budgets, as dependent on your organization's accounting policies. Its guidance on AI is the same: show teams their AI costs by project and workload before charging them back.

In practice, showback for a quarter lets department heads question the numbers and fix their own tagging before the cost hits their budget. Once the allocation is stable and the fallback is small, move to chargeback so AI spend sits in each department's actuals next to the budget it was planned in.

How to set up providers so the allocation runs itself

The allocation is only as good as the structure underneath it. The FinOps Foundation's AI guidance puts it simply: map each API key to one team, application and cost center. In practice:

• One workspace or project per team or application, never one shared key for everything.
• Separate production from experimentation, so customer usage lands in COGS and prototypes in R&D.
• Give every internal agent its own key, tagged to the function whose work it does.
• Assign seats through groups that match your HRIS departments, so the roster export maps cleanly.
• Buy through as few channels as possible. Corporate cards and expense reports hide AI subscriptions; pull 12 months of card and AP data once to find them.

What fallback weight should you use for untraceable AI spend?

The one that best matches who probably consumed it, written down and applied the same way every month. Three common choices:

• Share of traced usage. If 70% of tagged API usage came from engineering, assume 70% of the untagged usage did too. Best when most usage is already tagged.
• Engineering headcount. Reasonable when the untraceable usage is clearly technical, such as a shared development key.
• AI seat count by department. Reasonable when usage is spread across the company.

Whichever you choose, report the fallback amount separately and make shrinking it a goal. A fallback that grows month over month means new usage is being created without owners.

Common mistakes in allocating AI costs

• Booking the whole invoice to IT. It hides which teams drive the spend and makes every department's actuals look cheaper than they are.
• Spreading everything by headcount. Headcount is a fallback, not a driver; usage is almost never evenly distributed.
• Missing card and expense spend. Individual subscriptions bought on corporate cards are AI spend too.
• Allocating prepaid credits as they're bought. Allocate usage as it's consumed, and track the prepaid balance separately.
• Skipping the reconciliation. If allocated totals don't tie to the ledger, the board number and the department numbers won't match.

How to report the allocation to department heads

Give each department head a one-page view each month: their AI seats, their usage by workspace or project, their share of the fallback and the trend against last month. Show the allocation rules alongside the numbers so nobody has to ask how their figure was calculated. Department heads who can see their own usage tend to fix tagging problems faster than finance can, and they start asking the useful questions, such as whether a cheaper model would do. Our comparison of the best LLMs for finance teams covers current model pricing tiers, which is usually where usage costs come down first.

For the board, roll the same data up to total AI spend, its split across COGS, R&D and operating expense, and the largest departments, with one line on how much is still untraceable.

How allocation feeds the AI budget

Once each department has a few months of allocated actuals, the budget gets much easier. Seats follow hiring plans, usage follows customer growth or adoption, and each department can forecast its own line. Our guide to forecasting AI spend in next year's budget covers the drivers, the FY2027 budgeting software guide covers where the AI line fits in the wider plan, and the general method for any shared cost is in our guide to automating cost allocations. The From AI spend to AI budget replay walks through the whole process.

Allocate AI spend automatically with Aleph

Aleph connects your general ledger, HRIS and spend data to Excel and Google Sheets, so allocation rules are defined once, applied every month, and department AI actuals sit next to the budget automatically. Get the full framework in our budgeting for AI spend guide, see how Aleph's data consolidation works, or book a demo to see it on your own data.