AI cost outside the cloud bill is everything a company pays for AI that no AWS, Azure or Google Cloud billing export carries: seats in assistant and chat products, usage charged on top of those seats, model API invoices paid directly, AI features inside other SaaS, and purchases on a company or personal card that bypass procurement, often called shadow AI. Each of these sources has its own unit (seat, credit, token), its own document, and only sometimes names a person or a team. A FinOps scope drawn around a product that reads only cloud exports sees none of it.
How it works
Providers sell AI to employees in three shapes, which many plans combine; two further routes put AI cost in unexpected places.
A seat per person. A fixed monthly price per named user. Anthropic lists a Team plan standard seat at USD 20 a month billed annually or USD 25 billed monthly, and a premium seat with five times the usage at USD 100 or USD 125. GitHub lists Copilot Business at USD 19 and Copilot Enterprise at USD 39 per granted seat per month.
A seat with an allowance, and usage above it. At GitHub, each Copilot licence brings a monthly allowance of GitHub AI Credits, 1,900 on Business and 3,900 on Enterprise. One credit equals USD 0.01 and is consumed by model and token; code completions do not draw on it. The allowances of all licences form one pool for the billing entity, and whatever is left at month end is forfeited at 00:00 UTC on the first. Once the pool is empty, usage continues at the published rates and is charged on top. GitHub enables this additional usage by default; an administrator has to switch it off. Anthropic lets members on the Team plan and on seat-based Enterprise plans keep working past their seat's included usage with usage credits charged at API rates: bought in advance on Team, billed monthly in arrears on actual usage on seat-based Enterprise.
A seat without an allowance. On Anthropic's current usage-based Enterprise plan the seat fee buys access and no usage, listed at USD 20 per seat per month, billed annually, with a 20-seat minimum. All usage is a separate charge at API rates.
AI inside other SaaS. Microsoft bills agents in Microsoft Copilot Chat and SharePoint agents on pay-as-you-go through Azure meters on an Azure subscription. A billing policy in the Microsoft 365 admin center makes the link: it ties a subscription and a resource group to a group of users, and a tenant can hold up to 50 such policies.
Through a cloud invoice. Some of these purchases do end up on a cloud bill. GitHub can bill Copilot through a connected Azure subscription, sends usage to Azure daily and puts the calendar month on the Azure invoice at the start of the next. An Anthropic Enterprise plan bought through AWS Marketplace is paid through AWS, monthly in arrears on consumption. How such charges appear is covered in marketplace purchases on the cloud invoice; model API invoices paid directly are covered in AI provider billing.
How each source can be read:
| Source | Unit | Where the cost sits | Person or team | Read through |
|---|---|---|---|---|
| Claude Team, Enterprise | seat, usage at API rates in USD | Anthropic invoice, or through AWS | person and model | CSV export of usage analytics |
| GitHub Copilot Business, Enterprise | seat, AI credits | GitHub invoice or Azure invoice | person and model, cost centre | GitHub billing usage reports, the AI usage report among them |
| Microsoft Copilot Chat, pay-as-you-go | Azure meter | the billing policy's subscription | user group per policy | Azure cost data |
| Model API, direct | tokens | model provider invoice | project, workspace, API key | the provider's usage export |
| Card, expenses | subscription fee | card statement, expense report | cardholder only | statement file |
Anthropic's spend export holds one person and one model per row, with requests, tokens and cost before and after discounts, refreshed daily with a one-day delay. On seat-based Enterprise plans it shows only spend beyond the seat allowance. At GitHub, Copilot code review also consumes GitHub Actions minutes, attributed to the repository and from there to the cost centre: a different meter from the AI credits. A card statement names the merchant and the amount, but neither the team nor the purpose.
What it means for your cost
A complete view of AI cost, AI SaaS spend included, needs these sources beside the cloud data, converted into one currency and with a cost owner for each subscription. The list prices above are in USD, and Anthropic bills self-serve Enterprise plans in USD only; multi-cloud reporting in one currency covers bringing amounts into a reporting currency. Files without a running export reach the report through a file import.
A seat with an allowance moves where cost becomes visible. At GitHub, the allowance at USD 0.01 per credit is worth exactly the seat price: 1,900 × USD 0.01 = USD 19 on Business, 3,900 × USD 0.01 = USD 39 on Enterprise. A hundred Business licences make a pool of 190,000 credits, USD 1,900 a month, before any additional usage. Up to that point extra use costs nothing more; beyond it the bill rises with no seat added. Because the pool is shared, one heavy user draws down everyone else's allowance, so cost allocation to teams by seat count is not enough: it takes usage per person.
Three figures per tool belong in the monthly report: seats paid, seats in active use, usage above the allowance. Payment often lags or leads consumption. Anthropic charges Enterprise seats at the start of the annual cycle and seats added mid-term immediately, prorated for the rest of the year; on self-serve Enterprise, seats cannot be removed during the term. How to read such a payment spread across its term is shown in list, billed, effective and amortized cost.
Where it goes wrong
- Seats nobody uses. The invoice stays flat while usage analytics show fewer active people. The cause: seats are assigned but not reclaimed when people move or leave, or the contract allows no reduction before renewal, as on Anthropic's self-serve Enterprise plan.
- The same tool bought several times. Several teams pay for the same assistant in separate accounts on cards. The cause: without central purchasing each team opens its own account, with no shared pool, no admin export and its own invoice.
- API keys on a personal card. The spend shows up only as a reimbursement in an expense claim, weeks after the usage. The cause: the account with the model provider belongs to a person, not the company, and its projects, keys and exports leave with that person.
- Additional usage nobody decided on. The GitHub invoice exceeds seats times price. The cause: usage beyond the pool is allowed by default, and nobody set a budget or switched the policy off.
- An admin center budget read as a ceiling. Spend under a Copilot Chat billing policy passes its budget. The cause: Microsoft sends emails when the budget is reached but does not enforce it, and usage carries on.
- Counted twice. Copilot or a Claude plan appears twice in the report. The cause: the cost is already in the cloud export through Azure or AWS and was imported again from the provider's admin console.
Sources
Read on 2026-10-16. Anthropic Claude plan pricing, usage credits for Team and seat-based Enterprise plans, Enterprise plan billing and usage analytics for Team and Enterprise; GitHub Copilot plans, usage-based billing for organizations and enterprises, billing through Azure, models and pricing and billing reports reference; Microsoft Microsoft Copilot pay-as-you-go overview, setup and meters.