Your AI bill’s hidden gap is 40 to 60 percent of the total cost, because the invoice stops at the key rather than the team that spent the money. A model provider reports token use by the application programming interface (API) key, the project or the workspace — in the Foundation’s words, “at most by API key, project, or workspace, with no concept of business unit, cost center, or application” [1]. Everything else the feature needs is billed by someone else. The vector database, the rerankers, the agent loop and the log store all land on their own bills, under their own account names. So the clean split a finance team wants at month end is not a report sitting on any bill. It is a set of tags someone adds before the request leaves the app.

The big picture:

What I keep noticing is that the rush to split costs arrived before the tools to do it. The FinOps Foundation’s paper cites the State of FinOps 2026: 98 percent of respondents now manage AI spend, up from 31 percent two years earlier [1]. When this money sat in one test budget, nobody needed a map to find it. Now it is a real budget, and a line with no clear owner is the first thing a chief financial officer asks about.

Cloud platforms at least shipped a native tool. Cost tags let a team stamp spend with a cost center, and Amazon Web Services (AWS) warns that a tag can take up to 24 hours to reach the billing screen [2]. A token invoice carries no such field by default. Its own fields stop at the key, the project and the workspace, and an engineer types each one into a config file.

Some vendors now let a buyer push more detail in. Amazon Bedrock supports application inference profiles and cost tags that carry into cloud billing data, OpenAI offers project-scoped keys and a costs endpoint, and Anthropic reports usage by workspace [1]. That is real progress, and it stops well short of what a clean chargeback needs. None of those fields knows which unit owns the key, and none of them covers two vendors at once. The Foundation’s own ladder runs from account-level spend to per-workload cover to per-request unit metrics, with the split moving from nothing to showback to chargeback against team budgets [1].

Buyers have told the market where the gap sits. In the September 2026 State of Tokenomics survey, the most common request to model providers was more transparency and granular data, at 23 percent, while only 4 percent asked for lower prices [1]. Cost control became a strict rule for buying before it became a normal coding habit. This desk’s own record of that shift details why cost visibility turned into a buying requirement.

By the numbers

  • 40 to 60 percent — Off-invoice spend: In live retrieval-augmented generation (RAG) and agent setups, the harness — vector stores, embedding steps, rerankers, routing code, log capture — often carries 40 to 60 percent of total feature spend, and none of it shows on the model bill [1].
  • 98 percent — Spend management jump: The share of surveyed FinOps respondents who manage AI spend today, against 31 percent just two years earlier [1].
  • 30 to 60 days — Baseline wait time: Watch each app for that long, then set budgets at 110 to 120 percent of the baseline with alerts at 80 and 100 percent [1].
  • $0.10 — Metered log ingest: One vendor lists log ingest starting at $0.10 per ingested or scanned gigabyte a month [3]. That shows the ongoing cost of the monitoring leg the harness adds.

What I’d watch:

My sense is that the order of the work matters more than the software tools. The first chargeback argument always centres on the method. Here is what the people closest to this problem are doing.

  • One key, one owner: The cheapest control ties each key to one team, app and cost center, with a named person attached [1]. It gives rough numbers with no new tools, and it survives a company reorg.
  • Two streams, one label: The split is complete only when token spend caught at a proxy joins the harness spend from cloud bills on a shared app label [1]. A chargeback that counts just one stream is a partial bill with a confident number on it.
  • Showback before chargeback: Showback means showing a team its own number before anyone bills it. A team that has never seen its own AI costs cannot defend a sudden charge against its budget. The middle rung exists because the first fight is always about where the number came from.
  • The router as junction: 86 percent of respondents were trying or using a model router in 2026 [1]. A router sees the vendor, the model and the caller on one request, so it is the one place a split can be stamped without asking coders to remember anything.

The catch

My read, held loosely: this effort buys clear sight, not actual savings. A bill split perfectly and landing on the same profit and loss only changes who argues about the cost. The harness share is a working-group estimate rather than a strict measure, so a team that builds a plan on it should expect the first quarter to be wrong. Unit costs become a useful tool only once someone ties them to a business goal. The metering layer also carries its own bill, and a small program can easily spend more on watching AI usage than it splits.

At a glance

  • The Big Shift: Nearly every FinOps team now manages AI spend, but the invoice stops at the key or workspace, so a per-unit chargeback has to be built from scratch.
  • Why It Matters: A chargeback built from tokens alone bills each unit for less than half of what it uses, because the harness that makes the feature work lands on other bills.
  • What I’d Watch:
  • Key ownership: One key per team, app and cost center is the cheapest split that holds up.
  • Two streams: Token spend at the proxy plus harness spend from cloud bills, joined on one label.
  • The ladder: Showback first, and chargeback against team budgets only after the number survives one argument.
  • The Fine Print: The harness share is an estimate, unit costs need a goal attached before they help, and the monitoring layer itself runs up a bill.

Related reading

Sources

[1] FinOps Foundation — “Tokenomics: Managing AI Value in SaaS Model Token Costs” — https://www.finops.org/wg/token-economics-saas [2] AWS (Amazon Web Services) — “Organizing and tracking costs using AWS cost allocation tags” — https://docs.aws.amazon.com/awsaccountbilling/latest/aboutv2/cost-alloc-tags.html [3] Datadog — logs pricing, list price retrieved 2026-10-29 — https://www.datadoghq.com/pricing/