The bill nobody ordered
Ask an accounts payable team which supplier they find hardest to check this year, and many will name the same one: the AI provider. The invoice arrives in millions of tokens, split by model and by meter. Nobody in finance raised a purchase order for it. An engineer created an API key months ago, a team turned on a model in the cloud console, or someone put a subscription on a company card. Now the bill is real, it grows every month, and the only person who can say whether it is right is the person who ran up the usage.
In my last post I wrote that FinOps now manages AI spend, and that much of the data it needs already lives in procurement. This post is the practical side of that argument, for procurement and accounts payable. AI is a supplier like any other. It can be ordered, priced in a contract and checked before it is paid. It only needs a few checks that a normal invoice doesn't.
Four ways AI spend reaches you
The first difficulty is that AI spend doesn't arrive on one invoice. It comes through at least four routes, each billed in its own unit and each covered, if at all, by a different kind of order.
| Route | Examples | Billed in | What usually covers it |
|---|---|---|---|
| On the cloud bill | Amazon Bedrock, Vertex AI, Azure OpenAI, OCI Generative AI, models bought on a cloud marketplace | Tokens, model units, capacity hours | The cloud account's order, if there is one, which was sized before anyone used AI |
| Direct from the model provider | OpenAI, Anthropic, Mistral, OpenRouter | Tokens by model and meter, prepaid credits | Often nothing: a card, or an invoice approved because it looked like last month's |
| Data platforms and GPU clouds | Databricks, Snowflake, CoreWeave, Lambda, Nebius, Crusoe | Credits, GPU hours, reserved capacity | A contract with a commitment, rarely tracked against use |
| Seats | ChatGPT Enterprise, Claude Enterprise, Microsoft 365 Copilot, GitHub Copilot, Cursor | Seats a month | A renewal, with the seat count set a year ago |
Table 1. The four routes AI spend takes to accounts payable, and what usually covers each today.
The routes overlap. The same model can be billed by its maker and by a cloud that resells it, so the invoice to check is not always the one with the model maker's name on it. When Claude runs on Bedrock, AWS is the supplier that invoices, and Anthropic is only the model's vendor.
Why an AI invoice is hard to check
A normal invoice line says how many of something you bought and at what price, and you compare both with the order. An AI line says the same, but each half is harder to confirm:
- The unit depends on how the model was called. Input tokens, output tokens, cached input, batch requests, images and GPU hours each have their own price. Output usually costs several times what input does, and cached or batch tokens much less.
- The price moves. Providers cut list prices often, sometimes in the middle of a month. A line billed at last month's rate is not a typo you would spot by eye.
- The quantity can only be confirmed at the source. No one in finance counted the tokens. The provider's own usage report is the only independent record of how many were used.
- Seats are bought once and used less. A seat line is easy to read, but the question is not the price. It is whether those people still use the tool.
- Reserved capacity is paid for whether you use it or not. A reservation that ran at a third of its hours was billed correctly and still wasted money.
None of this is a reason to wave AI invoices through. It is a list of what to check.
Start with the purchase order
Every check below needs something to check against, and that is the purchase order. For AI, a good order line names four things: who invoices (OpenAI, or the AWS account), the model, the meter, and the amount approved for the period. Put the cost center and GL code on it, as on any order. If you have negotiated a rate, a discount off list price or a spend commitment, write it into the contract the order points to, as data and not only in the signed PDF.
The usual objection is that nobody knows the number in advance. That is true the first time, and it is why the order should follow the spend you can already see. In Pleaxy, the AI spend page lists what each provider has billed this month that no order covers, largest gap first, and drafts the purchase orders to cover it. The draft goes through the same approval chain as any order, so the first time a budget owner sees AI spend is not on an invoice they can't refuse.
Nine checks before you approve
With an order and a contract in place, each AI invoice line can be matched to its order line by invoicer, model and meter, never by its description, and held to these checks:
| Check | The line is compared with |
|---|---|
| Rate above contract | The rate the contract sets, or its discount off the list price, on the line's date. |
| Commitment exceeded | The overage rate the contract names for spend past its commitment. |
| Rate above list | The public list price for the period, where no contract term covers the model. |
| Price cut not applied | The new list price, when the provider cut it during the period. |
| Cache discount missing | The cached input rate, when cached tokens were billed at the uncached rate. |
| Batch discount missing | The batch rate, when batch tokens were billed at the interactive rate. |
| Tokens don't match usage | The provider's own usage report for the period, more than 3% apart. |
| Billed twice | Another invoicer billing the same model for the same project in the same period, such as a direct key and a marketplace offer. |
| Seats above active users | The users who signed in that month, from your identity provider or the provider's admin console. |
Table 2. The checks Pleaxy runs on every AI invoice line. Any finding puts the invoice in front of a person; none is approved automatically.
A tenth finding, reserved capacity used under its threshold, is not an error in the bill. It is flagged as a saving: a smaller reservation next time.
Figure 1 shows what this looks like on one invoice. The provider, the models and the prices are made up, but the checks are the ones above. Three of the four lines are questioned, for almost a quarter of the total, and none of the three would have stood out to someone reading the PDF.
Arcline AI · INV-20431 · October 2026
Matched to PO-8842
- Line 1arc-large · input tokens1,240 million tokens at $2.50 per million$3,100.00
Rate above contract: billed at $2.50 per million; the contract sets $2.25. $310.00 over.
- Line 2arc-large · cached input tokens860 million tokens at $2.50 per million$2,150.00
Cache discount missing: cached input billed at the uncached rate of $2.50; the cached rate is $0.25. $1,935.00 over.
- Line 3arc-large · output tokens410 million tokens at $10.00 per million$4,100.00
Matches the contract rate and the provider's usage report.
- Line 4arc-small · batch input tokens3,000 million tokens at $0.40 per million$1,200.00
Tokens don't match usage: billed 3,000 million; the provider's usage report shows 2,400 million, 20% apart.
- Billed
- $10,550.00
- Questioned before payment
- $2,485.00
Watch the month, not only the invoice
Checking the invoice catches a wrong bill. It doesn't stop a right one from being too big. AI usage can double in a week when a team ships a new feature, and the invoice that shows it arrives a month later.
So read the usage while it happens. With a read-only connection to each cloud account and AI provider, Pleaxy reads the AI lines every day and shows the month to date and a forecast beside what is left on the purchase order they will land on. Attribution rules turn the providers' tags, projects and workspaces into cost centers, so a cost center's budget counts its AI spend as it builds up, not when the invoice is approved.
The question a budget owner actually asks is short, and they can ask it as written. "Will this month's Bedrock forecast fit PO-8842?" The Pleaxy Assistant answers from the same figures, in Pleaxy, Slack or Teams, and so do Claude and other assistants through the check_ai_spend tool of the Pleaxy MCP server. The answer is a forecast, not a promise, and it never blocks anything on its own.
Put the limit where the spend happens
A purchase order is a decision about how much to spend. For most suppliers, finance enforces it at payment. For AI, by the time the invoice arrives, the money has been spent. The only place to enforce the order is at the provider.
If you choose to, Pleaxy sets the approved order line as your own spending ceiling at the provider: an OpenAI project limit, an Anthropic workspace limit, a Google Cloud budget or spend cap, or an AWS budget action. It uses a separate credential you create for the purpose, never the read-only one, and you can remove the ceiling at any time. Be clear about what each provider can do. Some stop spend at the ceiling. AWS budgets refresh a few times a day, so a stop can come hours late. Azure budgets only alert.
A checklist for procurement and AP
If AI spend reaches your desk today without an order, these six steps bring it under the same rules as everything else:
- Find every route. List the cloud accounts with AI services turned on, the providers with an API key, the platforms with credits and the tools billed per seat. Card statements are where the surprises are.
- Raise an order for each invoicer. Size it from the last three months, name the model and the meter, and put it on a cost center.
- Write the rates into the contract. A negotiated rate, a discount off list price or a commitment can only be checked if it is recorded as data.
- Ask for usage, not just the invoice. Get read-only access to each provider's usage report, so the quantity on the invoice can be confirmed at the source.
- Count seats every month. Compare seats billed with users who signed in, and bring the number to the renewal.
- Decide where to stop, not only where to alert. Agree with each budget owner which orders should be a hard ceiling at the provider and which only a warning.
How Pleaxy fits, and its limits
Everything above is part of Pleaxy's AI and cloud spend add-on. I work for Pleaxy, so I want to be just as plain about what it doesn't do:
- Pleaxy reads usage only where a provider has a billing API it can read with a read-only key: the large clouds, OpenAI, Anthropic, Mistral (where the organization has its Admin API), OpenRouter, CoreWeave, Databricks and Snowflake. Lambda, Nebius, Crusoe and the seat-based tools have none, so their invoices are held to the contract and the order, but not to a usage report.
- Pleaxy doesn't read your identity provider. You send it each month's active users for a seat product, from a script, an agent or by hand.
- On AWS, AI lines come only through Cost Explorer, so the connection needs access to current costs for any AI figure.
- Amounts stay in each provider's billing currency. Pleaxy never converts them or adds up different currencies.
- The connection to a provider is read-only and never changes anything there. Spending ceilings are opt-in and use their own credential.
If you would like to see your own AI spend against your orders, talk to us.
About the author
Andrea Gneri
Sales leader, Pleaxy
Andrea Gneri leads sales at Pleaxy. She works with finance, procurement and FinOps teams on how to bring their suppliers, contracts and invoices into one place, and writes about what she learns from them.