FinOps for AI: what changes when the bill is tokens.

01What FinOps for AI means.

FinOps for AI means managing what AI costs and what it returns with the same shared discipline FinOps brought to cloud, where engineering and finance read one view of spend and make trade-offs together. AI FinOps is the same thing under a shorter name.

It has become the default job of a FinOps team in two years. In the FinOps Foundation's State of FinOps 2026 survey of 1,192 practitioners, 98 percent said they manage AI spend, up from 31 percent two years earlier. The same survey named FinOps for AI as the top forward-looking priority.

The scope covers AI bought as seats in chat and coding tools, as usage through model APIs, and as agents that call models on their own. The reverse phrase, AI for FinOps, means using AI to do FinOps work, and the Foundation treats it as a separate topic. Before any of the practice below, you need the total across every route. Our AI spend management guide covers how to find it.

TakeawayFinOps for AI applies the cloud FinOps practice to seats, API usage and agents, and nearly every FinOps team now does it.

02How AI FinOps differs from cloud FinOps.

Cloud FinOps manages resources that run for hours and carry tags. AI FinOps manages calls that last seconds and are priced per token, often made by an agent acting on its own. The FinOps Foundation's FinOps for AI overview sets out the differences.

• The unit. Bills count tokens, API calls and provisioned throughput. The tokens billed are often more than the tokens a user types, because the application adds instructions, history and retrieved content to every call.
• The price. The Foundation notes that models are released and repriced often, and that prices can move sharply in either direction. Output tokens often cost three to five times as much as input tokens.
• The tags. New SKUs and direct provider bills often lack native tagging, so attribution needs instrumentation in your own code or in the path each call takes.
• The spender. An agent decides how many calls a task takes. Gartner expects AI inference costs per agentic workflow to increase more than fivefold through 2028, even as the price of each token falls.
• The forecast. Usage is harder to predict, especially early, and the cost of the same task varies from one run to the next.

TakeawayThe unit is the call and agents choose their own volume, so the cloud playbook needs instrumentation at the call.

03How the FinOps Framework maps to AI.

The FinOps Framework is built on three phases, inform, optimise and operate, with capabilities grouped into four domains, and it lists AI as one of the technology categories it covers. Most capabilities keep their name for AI and change in practice.

Teams cycle through the phases continuously. Inform shows cost, usage and value. Optimise lists the options to improve them. Operate puts the chosen changes into effect, and the loop starts again. The table maps the capabilities that change most for AI to what each means in practice, and to the guide that covers it.

TakeawayKeep the framework and change the unit: each capability now works at the level of the call and the use case.

04How to attribute AI costs to teams and agents.

Attribute each model call to a team, a project and, where one is involved, an agent, at the moment it runs. Providers bill by key and project, so the owner has to be recorded in the path the call takes, before the invoice arrives.

1. Give each team and each production agent its own key or project, so the provider bill already splits along owners.
2. Tag every call with team, project, agent and environment. The Foundation recommends tagging training separately from inference, and production separately from experimentation.
3. Record the cost when the call runs, at that day's price.
4. Give untagged spend a named owner, and report its share every month until it is close to zero.

Multi-agent work is the hard case. The Foundation notes there is no accepted method yet for allocating it, so write your rule down before the first report, such as charging the agent that started the task. Our guide to tracking LLM costs per team covers the tags, the ledger and a four-week rollout.

Then report it back. The Foundation's advice is to report through showback first, showing teams their AI costs by project and workload before charging them back. Chargeback makes sense once a team's spend is material and its numbers are trusted. Our showback and chargeback guide sets out when to use each.

TakeawayTag the call when it runs, and show teams their spend before you charge them for it.

05How to forecast AI usage.

Forecast AI spend from its drivers, as a range, over shorter windows than cloud. Volume, calls per task and price each move on their own, so a forecast built from last month's invoice misses the change that matters.

• Volume: users, tickets, documents or tasks per month for each use case.
• Calls per task: an agent can make many model calls for one request, and context grows over a session because most APIs resend the history each time.
• Tokens per call, split into input and output, since output is priced higher.
• Price per model, checked against the provider's current rates, with room for repricing.
• Adoption: new users, new teams and new agents moving from pilot to production.

The Foundation recommends shorter windows and more frequent revisions for AI forecasts. It also advises reporting cost per run as median and high-percentile figures, because a single average hides the expensive runs. Re-forecast monthly, and whenever a model, prompt or reasoning setting changes.

TakeawayForecast from drivers as a range, and re-forecast whenever a model or prompt changes.

06How to set AI budgets that hold before the spend.

Set budgets in dollars, per team and per agent, and enforce them where the call happens so they act before the money is spent. A budget read after the month closes is a record of what happened.

The Foundation's overview says top-down budgets are less accurate for AI and that pilots need flexible funding. Its guidance on tools goes further, calling budget controls and maximum iteration limits essential cost controls. Most large companies are partway there: in KPMG's Q3 2026 AI Pulse survey of 314 US leaders at companies with revenue of US$1 billion or more, 43 percent had implemented usage or token budgets.

1. An organisation ceiling for the year, released month by month.
2. An allocation for each team, drawn from the ceiling.
3. A cap for each agent and each person, counted across every provider.
4. A token or step limit per task, so a loop stops early.
5. A named approver for every increase, with a reason and an expiry.

Our guide to token budgets and spending limits for AI agents covers what should happen at each cap and who approves an increase.

Each provider's own cap stops at that provider's edge, so treat native caps as a backstop under one budget. Our comparison of AI spending limits shows what each provider lets you cap.

TakeawayEnforce dollar budgets per team and per agent in the path of the call, with provider caps underneath.

07How to respond to AI cost anomalies.

Treat an AI cost anomaly as an incident: detect it within minutes and slow or stop the agent behind it while you find the change that caused it. The Foundation expects AI anomalies to be more frequent and riskier than cloud ones, and normal ranges harder to define.

The usual causes are agent loops, prompt changes that raise tokens per call, model switches that change the price, and runaway training or fine-tuning jobs. The Foundation advises monitoring tokens per minute at the application level, as well as daily spend. A daily total finds a loop the next morning. A rate alert finds it while it runs.

1. Alert the owner of the agent or application, named in advance.
2. Apply the default set for that cap: stop, slow, step down to a cheaper model, or wait for approval.
3. Find the change behind it, such as a deploy, a prompt edit, a new model or a traffic spike.
4. Record what the anomaly cost, and report anomaly cost as a share of spend each month.

TakeawayAlert on the token rate, and act on a default agreed before the anomaly happens.

08How to measure AI unit economics and value.

Measure cost per successful outcome for each use case, such as cost per resolved ticket, and set it against the value of that outcome. The Foundation recommends tracking cost per inference, per user and per successful outcome, reported as distributions.

Its overview lists the measures a FinOps for AI practice reports.

• Cost per inference.
• Cost per million tokens.
• Cost per successful outcome.
• Resource utilisation efficiency.
• Cache hit rate.
• Model fit: the share of requests sent to a more expensive model than the task needs.
• Anomaly cost as a percentage of spend.

Cost per token can fall while the bill rises. Gartner calls this the inference paradox, in which better unit economics raise the overall cost of AI without a clear path to matching value. Few organisations close the loop today: KPMG's global Q3 2026 AI Pulse found that 12 percent consistently assess AI value against cost. Our ROI guide gives finance a method, with a worked example.

TakeawayReport cost per successful outcome beside the value of the outcome, because a cheaper token can still mean a costlier task.

09What to look for in AI FinOps tools.

An AI FinOps tool should record the cost of every call with its team, application, agent and model, and act on a budget before the spend. The FinOps Foundation's guidance on FinOps for AI tools and services gives a checklist, summarised here.

• Per-call data: team, application, model, input and output tokens and cost on every call.
• Agent tracing: how many tokens a single agent run consumed, across all its model calls.
• Allocation: showback and chargeback by team, product or business unit.
• Budgets: thresholds that alert or stop when a team or application reaches its limit, and iteration limits for agents.
• Anomaly alerts on spend.
• Discovery: spend from direct subscriptions, cloud marketplaces and expense claims, as well as provider APIs.
• Unit economics: cost per use case, beyond total spend.
• Every provider in one view, so one budget can cover them all.

Our guide to choosing an AI cost management platform turns these into questions to ask in a demo.

TakeawayChoose a tool that sees every call with its owner and can stop spend at a budget, across every provider you use.

10Common mistakes with FinOps for AI.

• Copying the monthly cloud cadence. AI spend moves by the minute, so a monthly report finds a runaway agent a month late.
• Budgeting in tokens. Every model prices tokens differently, so set budgets in dollars and track tokens as a driver.
• Sharing one key across teams. The bill cannot be split afterwards, and nobody owns the number.
• Charging back before attribution is trusted. The first month of disputed bills costs more goodwill than it saves.
• Alerts with no stop. An alert that arrives after the money is spent describes the overrun.
• Forecasting from the invoice. Forecast from volume, calls per task and price, and revise monthly.
• Measuring cost per token alone. A cheaper model that fails more often can raise the cost per outcome.

TakeawayBudget in dollars at the level of the call, and stop spend before it happens.