FinOps for AI: Controlling Agent Costs in 2027

There is a new line item on your invoice that nobody approved and nobody owns. Last year it was a rounding error. This year it has a comma in it, and next year, once a handful of agents move from pilot to production, it will have two. The uncomfortable part is not the size of the number. It is that most companies cannot say which team, feature or workflow produced it, which means they cannot cut it without guessing.

That is the gap FinOps for AI closes. FinOps is the operating discipline that made cloud spend legible a decade ago: give every cost an owner, report it where decisions get made, and set guardrails before the bill surprises you. The FinOps Foundation's 2026 survey found that almost every team now manages AI spend, yet the capability they most wanted and did not have was granular monitoring of tokens, model calls and GPU time. In other words, people are on the hook for a number they cannot see. This post is about getting the number into focus and back under control before 2027 budgets lock in.

Why your cloud cost playbook breaks on AI

The instinct is to treat AI like another cloud service and point your existing cost tools at it. It does not fit, for three reasons.

• The unit is a token, not an hour. A server costs the same whether it is busy or idle. An agent costs nothing when nobody calls it and a surprising amount when someone wires it into a loop. Spend tracks usage, and usage is spiky and hard to forecast from last month.
• One task hides many calls. A single user request can fan out into a dozen model calls, tool calls and retries. If you only meter the top-level call, your attribution is wrong by an order of magnitude. Cost lives in the full request tree, not the first line of it.
• Nobody holds the budget. Cloud spend grew up inside engineering, with owners and tags. AI spend leaks in from everywhere: a marketing team's copy tool, a support team's assistant, an analyst's script. The invoice is central; the decisions are not.

We went deep on where the money actually goes in AI agent cost optimization and on the full bill in the total cost of owning an AI agent. FinOps is the layer above both: not how to make one agent cheaper, but how to run all of them with the lights on.

Start with visibility, because you cannot govern a bill you cannot read

Before allocation, before guardrails, you need to see spend at the grain decisions are made. That means tagging every call with who asked for it: team, application, feature, environment, and ideally the end customer. The target is to answer "what did the invoicing agent cost us last week, and per processed invoice" without a spreadsheet archaeology project.

The practical move is to route traffic through a single choke point so the tagging happens once, consistently, instead of being reinvented in every codebase. That is one of the strongest arguments for an AI gateway: it is where you meter tokens, attach cost attribution and enforce limits for every model and every team at once.

Showback before chargeback

Once costs are attributed, you decide how hard to push them onto teams. There are two levers, and the order matters.

Showback reports each team's AI spend to that team without moving money. It builds awareness and surfaces the workflow quietly burning 40% of the bill, and it does it without a political fight.

Chargeback books those costs against the team's actual budget. It drives real accountability, because a number that hits your own budget gets managed in a way a report never does. But it demands clean attribution. Charge a team for spend you tagged wrong and you lose their trust for a year.

Start with showback to earn trust and prove the numbers, then move to chargeback once attribution is solid. A common middle path works well: chargeback the 70 to 80% of spend you can attribute cleanly, and showback the shared remainder rather than forcing precision you do not have. Shared costs, a common retrieval index or a platform team's tooling, are fine to socialise rather than split to the cent.

Guardrails: budgets that act, not budgets that watch

Reporting tells you what already happened. Guardrails stop the bad month before it lands. The ones that pay for themselves:

1. Hard budget caps per team and per app, enforced at the gateway, that throttle or alert when a workflow runs away. A runaway loop should trip a limit in minutes, not show up on next month's invoice.
2. Model routing by job. Not every task needs the biggest model. Routing the easy 80% to a smaller, cheaper model and reserving the frontier model for the hard 20% is the single largest lever most teams have, and it ties directly to a multi-model strategy.
3. Caching and retry limits. Repeated identical calls and uncapped retries are pure waste. Cap them.
4. Access control at the gateway. Spend and security converge here: the same choke point that meters cost should govern which agents can call which models and tools, the case we make in the MCP gateway.

Measure cost per outcome, not cost per token

The last shift is the one that changes decisions. Cost per token is an engineering metric; it tells you nothing about whether the spend was worth it. The number that matters to the business is cost per outcome: per resolved ticket, per processed invoice, per qualified lead, per closed deal.

Framed that way, a model that costs twice as much per token can be the cheaper choice if it resolves the task first time instead of bouncing to a human. This is exactly why we argue to judge AI on resolution, not deflection, and why the how to measure AI ROI question has to start from the outcome. An agent that is cheap per call but hands half its work to a person is not cheap. It is a subsidy with good PR.

Once you can express spend as cost per outcome, the 2027 budget conversation stops being "how much will AI cost" and becomes "which workflows earn their keep." That is the question a CFO can actually act on, and it is the frame we use in what survives the 2027 AI budget.

What to do this quarter

You do not need a FinOps team to start. You need three moves before the year closes. First, pick one production agent and instrument its full request tree so you have one number you trust completely. Second, turn on showback for the teams using AI, even a rough monthly email, so the conversation starts from data. Third, put one hard budget cap in place at whatever choke point you have, so a runaway workflow trips a limit instead of a quarterly review.

The companies that scale agents through 2027 without a budget crisis will not be the ones that spent the least. They will be the ones who could always answer what they were spending and on what. If you are wiring agents into production and want the cost controls built in from the start rather than bolted on after the first scary invoice, tell us how your stack is set up and we will help you put the meter where it belongs.