What the Bill Cannot See

It is budget season. Somewhere in your organization, someone is building next year's AI number, and they will bring you a line item most budgets never carried before agents.

It is the consumption line. Credits spent by agents that ran overnight. A reconciliation workflow that loops until it finishes. A research agent that fanned out to twelve sources before it answered. The invoice is precise about every one of them. It can tell you which agent spent what, on which model, in which week.

Then someone asks the question that actually decides the budget. What did we get for it?

The invoice goes quiet. Not because the data is wrong, but because that answer was never on the invoice. It lives in the workflow: whether the reconciliation closed, whether a person had to redo the research, whether the agent finished the job or handed it back half done.

That gap is the subject of this quarter. Through Q3, Token Exchange argued that the token is becoming the native unit of work: an asset you can hold on a balance sheet, a cost you have to classify, and a currency whose exchange rate you need to read. September closed on what belongs in the quarterly business review. Q4 answers it with the harder follow up. If the token is the unit of work, how do you measure the work?

The market has just made the first half of that answer much easier.

Between June and September 2026, every major cloud and workplace platform shipped spend controls for agentic work. Cost agents that investigate anomalies and run recurring FinOps workflows. Hard monthly caps, pooled quotas, and anomaly detection for agent workloads. Per-seat licensing paired with usage-based billing for agents, with budgets and guardrails attached. Practitioners got there first: the FinOps Foundation's State of FinOps 2026 found that 98 percent of respondents now manage AI spend, up from 31 percent two years earlier.

This is a category shift, not a product launch. It brings AI consumption under guardrails, allocation, and reporting that finance teams already understand. For anyone who spent Q3 arguing that token spend needs to be read and classified like any other material cost, this is the right move, made at scale.

Look closely at what the releases promise alongside control, and a second word keeps appearing: value. Measure it, show it, prove ROI to leadership. Control and optimization are what FinOps has always done for cloud. Value is a commitment to something the bill alone cannot supply. The spend side of that promise arrives from the meter. The value side has to arrive from somewhere else.

September used McKinsey QuantumBlack's analysis of agent economics to answer an objection. If tokens are only a quarter of an agent's variable cost, why govern the token? Because it is the one line nobody outside the vendor can check. October turns the same numbers around. If the token is the quarter you govern, what governs the other three?

Oversight is not unaccounted for. Expert hours are salaries, and finance knows exactly what they cost. What finance cannot see is which agent, which workflow, and which completed task those hours belong to. The cost is on the books. The attribution is not.

McKinsey draws the operational conclusion directly. Many organizations, they note, focus optimization on model selection and token costs because those are the most visible expenses, when the larger gain comes from redesigning workflows to reduce exceptions and simplify review. The cheapest token is not the lever. The review queue is.

This is not an argument against the token as the unit of work. It is the reason the unit needs a denominator. Spend per token tells you the price of effort. Only completed work tells you the price of an outcome.

McKinsey gives the missing half a name. Just as organizations built FinOps to manage cloud spend, they argue, business units now need an equivalent discipline for continually managing agents.

The word to notice is equivalent, not replacement. AgentOps does not compete with FinOps for AI. It feeds it. FinOps owns the cost of effort. AgentOps owns the evidence of outcome. Put them together and outcome per dollar spent on AI collaboration stops being an estimate and becomes a calculation.

The last row is the one that should get a CFO's attention. FinOps is very good at catching the agent that costs too much. It has no way to see the agent that costs exactly what was forecast and hands its work back to a person every time. That agent looks healthy on any spend dashboard, because on spend it is. Only the workflow knows.

McKinsey is candid that the hard questions are still open. Three of them land squarely on the people building next year's budget.

The cost coding gap is the one September found in vendor credits, one layer up: a number on the books that nobody can trace to the work it paid for. None of these are FinOps failures. They are the edge of what a billing system can know. Crossing that edge is an operating problem, which is why it needs an operating discipline.

Until that discipline exists, one test separates a line you can defend from a line you are hoping about. Run it on every agent line in next year's request.

Three yes answers and the line is defensible. Anything less is not a line to cut. It is a line to instrument, which is what the moves below are for.

The test tells you which lines fail. These moves fix them, and none of them require new tooling. They require deciding, before the number is locked, what the number is supposed to buy.

• Name the outcome for every agent line. For each consumption line in next year's request, write down the completed task it pays for: a closed reconciliation, a resolved ticket, an accepted draft. A line with no named outcome is a line nobody will be able to defend in Q2.

  Building a personal impact report across four AI platforms, the first step was not counting sessions. It was naming the job each workflow was hired to do. The workflows that resisted a name were the same ones that resisted a value.

• Count acceptance, not activity. Runs, sessions, and tokens measure effort. Track how many outputs were accepted without being reopened or redone. That single ratio exposes the agent that is on budget and producing nothing.

  In my own instrumented report, the sessions that needed the most scrutiny were the ones with no clear outcome. The method counts their time and flags them rather than folding them into value. An honest receipt includes the part you cannot yet explain.

• Put the review queue on the same ledger. Estimate the expert hours spent checking each agent's work and attach them to the same workflow as its token spend. If McKinsey's pattern holds in your environment, this is where most of the variable cost has been hiding.

  Running the same report across platforms surfaced one quiet error: the same person's time valued at different hourly rates in different reports. Oversight cannot be compared until it carries one rate on one ledger.

• Give one workflow a cost per outcome by Q1. Pick a single high-volume workflow and derive its fully loaded cost per completed task: tokens, oversight, and fixed run cost, divided by accepted outputs. One derived number beats a portfolio of estimates in a QBR.

  Of the four platforms I instrumented, only the one that meters per task let the cost side be derived rather than estimated. The value side still rested on estimated time. One workflow with both sides derived would outweigh all four reports.

• Assign the AgentOps owner before the FinOps rollout finishes. Spend controls are arriving now. Decide who owns the outcome side of each workflow while the cost side is being configured, so the two land joined rather than reconciled later.

  Across every report, the value pillar that stayed nearly empty was risk mitigation. Not because the work carried no risk value, but because nothing in the workflow was assigned to look for it. Outcomes nobody owns do not get counted.

FinOps for AI is half of an instrument. The organizations that pull ahead will build the other half before the budget closes.

The other half is operational. It lives in the review queue, the exception rate, and the count of work that was finished and stayed finished. Join the two and the token stops being a cost to contain. It becomes a unit of work with a price you can read on both sides of the ledger: derivable on the invoice, attributable on the outcome.

That is the measurement this quarter is about, and it starts with one number. Token Exchange already publishes an instrument for half of it: What My Agents Did for Me, an open-source impact report that runs across four AI platforms, counts sessions from records, and maps each workflow to the value it was hired to create. Today it derives the work and estimates the hours. November closes that gap: one workflow, cost derived from metered records, outcome derived from accepted work, joined on the same task.

The budget question in AI is no longer what you spent. It is what the spend finished.

FinOps reads the invoice. AgentOps reads the work. Only together do they read the return.

Token Exchange · The market signal for the agent-powered economy
© 2026 · kadns.ai