FinOps for AI: budgets, forecasting & the cost review

TL;DR

Cost attribution tells you where the money went, after it's gone. FinOps is the operating discipline that decides where it goes next — before the invoice, not after. Four practices make it real: budgets and quotas enforced in code, not just watched on a dashboard; chargeback or showback so a team feels the cost of what it builds, instead of every AI feature drawing from one undifferentiated pool; a forecast that accounts for two curves moving in opposite directions at once — usage climbing, token prices falling — and a naive extrapolation of last month's bill gets both wrong; and a recurring cost review, structured like a postmortem, that catches drift before it becomes a headline number on a board slide.

🎯 For the product leader

Why it matters — Attribution without governance is a very detailed receipt. FinOps is what turns "we can see where the money went" into "we decided where it should go."

What it changes in your decisions — Budgets get enforced, not just alerted on. Teams see their own AI spend, not a shared line item nobody owns. Forecasts account for both usage growth and falling per-token prices instead of naively projecting one line.

Ask your eng team — "If a feature's spend doubled overnight, would anything stop it automatically, or would we find out from the invoice?"

Risk if ignored — A quietly runaway feature burns budget for weeks before anyone notices, or a forecast built on last month's per-token price is wrong by the time this quarter's board deck ships.

From attribution to governance

Budgets and quotas — enforced, not watched

A dashboard that shows spend rising is attribution. A budget is attribution with teeth: a per-tenant or per-feature cap that does something automatically when it's hit — throttles, degrades to a cheaper model, or pages a human — rather than a number a person has to notice and act on manually. The distinction matters because the failure mode of a watched-only budget is specific and common: a runaway agent loop, a cache-busting change, or a viral feature can burn a month's allocated spend in a day, and a dashboard nobody was staring at in real time catches it only after the money is gone.

• Hard caps for anything where "run out and stop" is safe — a free tier, an internal tool, a feature where degrading gracefully costs nothing.
• Soft budgets with escalation for anything where stopping outright breaks a paying customer's workflow — alert a human at 80%, and again at 100%, rather than cutting them off mid-task.
• Caps inside the request. Set a per-request ceiling on output. OpenAI's max_output_tokens limits reasoning and visible tokens together. Anthropic reports thinking tokens in usage.output_tokens_details.thinking_tokens, so you can track them. For agents, also cap the number of turns and tool calls per task.
• Per-tenant and per-feature, not just aggregate — an aggregate budget can be healthy while one tenant is quietly destroying margin underneath it, invisible until someone slices the number the way cost attribution already knows how to.

Unit economics — cost per outcome, not per token

A token count tells you little on its own. The FinOps Foundation calls tokenomics "FinOps applied to AI". It extends unit economics to a resource that is priced per inference. Its guidance is to track token quality as well as token volume, because volume alone misattributes cost. Pick a unit your business already understands: cost per resolved ticket, per completed workflow or per active user. Then compare it to the revenue or savings that unit brings. A feature whose cost per resolved task rises while its price stays flat is losing margin, even if the token price fell.

Chargeback and showback — who feels the cost

Both practices route attributed cost back to the team that generated it. They differ in what happens next. Chargeback actually debits the owning team's budget — the cost is real money out of their P&L, and it changes incentives immediately: a team that pays for its own token spend starts asking whether every feature needs the biggest model. Showback only makes the number visible, with no financial consequence — cheaper to implement, and often the right first step, but it relies on teams caring about a number that doesn't touch their budget. The tradeoff is real: chargeback creates sharper incentives and more friction (billing disputes, contested attribution when a shared service's cost gets split); showback creates less friction and weaker incentives. Most organizations start with showback to build the attribution muscle, then move to chargeback once the data is trusted enough that a team won't dispute the number it's being charged.

Forecasting two curves that move in opposite directions

A naive AI cost forecast takes current spend and multiplies by expected user growth. It's wrong, because several forces move at once and in different directions. Write the forecast as a product: users × tasks per user × tokens per task × price per token. Each factor gets its own assumption. The first two are usage. The last two are the ones teams forget.

• Usage grows — more users, more requests per user as a feature matures, longer average context as conversations and documents grow. This pushes cost up.
• Per-token prices fall — model providers have cut prices steeply and repeatedly since the category began, and a well-run serving stack's marginal cost per token falls further as batching and paged attention mature. This pushes cost down.
• Tokens per task rise — reasoning models bill thinking tokens as output tokens, and agents make many model calls per task, each one re-sending context. Gartner's March 2026 forecast puts agentic tasks at 5 to 30 times the tokens of a chatbot exchange. (Source: search result only, not opened.) A falling price per token can sit beside a rising bill.
• The model mix shifts — a team that moves to a newer or higher tier pays that tier's price. A tokenizer change can also raise the count for the same text, as Anthropic reports for Claude 4.7 and later.

A forecast that only tracks usage overshoots, sometimes badly enough to justify a self-hosting decision or a price increase that a more honest number wouldn't support. A forecast that only tracks falling prices undershoots, and a team gets blindsided when a successful launch's volume outpaces the price relief it was counting on. The discipline is modeling each factor explicitly, with their own separate assumptions, rather than collapsing them into one trend line drawn through last month's invoice.

The cost review — a recurring ritual, not a fire drill

The practice that makes the other three durable: a standing, cross-functional review — engineering, product, finance — on a fixed cadence, structured the way technical product management structures an incident postmortem. Look at spend against budget per tenant and feature, check whether the forecast's two curves are tracking reality, revisit the build-vs-buy breakeven if volume has moved meaningfully, and — the step that's easiest to skip — assign an owner and a date to whatever the review finds, the same discipline that keeps a postmortem's action items from dying in a document nobody reopens. A cost review that only produces a slide with numbers on it, and no owned action items, is theater with a spreadsheet attached.

A worked pass: the forecast that was wrong in both directions

A team forecasts next year's AI infrastructure cost by taking this month's per-user spend and multiplying by projected user growth — a single line, no separate treatment of price and volume. Actual costs come in lower than forecast for the first two quarters, because the model provider cut prices twice in that window, and the team congratulates itself on efficiency gains it didn't actually make. In quarter three, a popular new feature pushes average conversation length up sharply, volume grows faster than the original forecast assumed, and per-token prices stop falling as fast — the same single line now understates cost, and a budget conversation that should have started in quarter two happens under pressure in quarter three instead. Both misses trace to the same root cause: one line trying to represent two curves that move independently. A forecast that modeled usage growth and price decline as separate assumptions would have caught the shift in either direction with two quarters of runway, instead of being surprised twice by the same conflated number.

Failure modes

• Watched, not enforced — a budget dashboard exists, but nothing automatic happens when a number crosses it, so the catch depends on someone noticing in time.
• Showback with no teeth, forever — a team never feels its own spend, so no incentive ever changes, and showback becomes a report nobody acts on.
• Price-per-token blindness — the team tracks the price of a token and misses that tokens per task doubled after a switch to a reasoning model or an agent.
• One-line forecasting — usage growth and price decline get collapsed into a single trend, producing a number that's confidently wrong in whichever direction the two curves diverge fastest.
• The cost review that never happens twice — a review gets run once, produces a good slide, and never recurs, so the next quarter's drift goes uncaught until the invoice.
• Action items with no owner — the review correctly identifies a problem, and nothing changes, because no one was assigned to fix it by a date.

Practitioner checklist

• Do our budgets actually do something automatic when breached, or only alert a human who might not be watching?
• Can each team see — or better, feel — its own AI spend, not just an aggregate number nobody owns?
• Does our forecast model usage, tokens per task and price per token separately, or one naive extrapolation?
• Do we report cost per outcome (per resolved task, per workflow), not only per token?
• Is there a recurring, cross-functional cost review on the calendar — and does it produce owned action items, not just a slide?
• When did we last revisit the build-vs-buy breakeven against current volume and current prices?

Related lessons

• The cost stack, and the build-vs-buy breakeven
• Cost attribution
• Incidents & postmortems — the ritual structure this lesson's cost review borrows.
• The economics of infrastructure

Sources

Checked 2026-10-29. FinOps Foundation: token economics, OpenAI reasoning guide, Anthropic extended thinking and Anthropic pricing.