You Are Buying Tokens. You Should Be Buying Finished Work.

Ask a CFO what the company spent on AI last quarter and you will get a number. Ask what that spend produced and you will usually get a pause.

That pause is the problem. For two years, leadership told teams to adopt AI, and usage became the scoreboard. Tokens went up. Licenses piled up. Some companies ranked employees on internal leaderboards by how much AI they consumed. Very few can now point to the specific spend that produced work that would otherwise have cost more, taken longer, or not shipped at all.

It is not a model problem. It is an operating problem, and it responds to the same discipline we have always applied to IT, vendor, and labor spend.

Volume Was Always the Wrong Metric

AI pricing rewards volume. For two years, volume was also the internal measure of success. Put those together and bills get away from people.

Uber is the public example. The company used up its entire 2026 AI budget in the first four months of the year, with about 95 percent of its engineers using AI tools every month. Its president then said publicly that the link between all that usage and better products for customers was "not there yet."

Uber made the news because the numbers were large. The common version is quieter. One team ships a chatbot. Another wires a coding assistant to the largest model available. A third runs the same retrieval job every hour because nobody owns the workflow. Finance sees a rising line item and cannot tie it to a process, a baseline, or an accountable person.

Spend first, measure later. That is how most AI estates were built.

A newer model does not fix this. On many common enterprise tasks, the leading models now perform close to one another. The advantage sits in what you build on top: routing, workflow design, and the discipline for deciding which work goes where.

What a Token Bill Hides

A token count measures activity. It says nothing about whether any work got done.

One line on the invoice can hold a frontier model doing classification a small model handles fine, the same long context re-sent on every call because nobody turned on caching, retry loops burning tokens on failures, and scheduled jobs that stopped being useful months ago. None of it shows up in a monthly total.

The larger cost usually is not on the AI invoice at all. It sits in the human rework that follows weak output.

The numbers below are illustrative. The structure of the math is what matters.

Consider an insurer running a claims summarization workflow. It makes 50,000 calls a month at $0.40 each, so the token bill is $20,000. Adjusters reject and rewrite 30 percent of the summaries. Each rewrite takes ten minutes at a loaded cost of $60 an hour. That is 2,500 hours and $150,000 of rework every month.

Look at what the insurer actually bought. It paid for 50,000 summaries and received 35,000 finished ones. The other 15,000 were drafts a person had to finish. The invoice says each summary costs $0.40. The real cost is $3.40, more than eight times higher.

Now compare two fixes. Switching to a model at half the price saves $10,000 a month. If quality slips and rework rises to 35 percent, it adds $25,000 in labor, and the company is $15,000 a month worse off while reporting a lower AI bill. Fixing the prompt, the context, and the validation step so rework falls to 10 percent saves $100,000 a month and brings the real cost per summary to $1.40, without touching the vendor.

The token bill was the small number. Teams that manage only what shows up on the invoice will keep pulling the wrong lever.

Finished Work Per Dollar Is the Scoreboard

Vendors price in tokens, and that is not going to change. What you can change is what you measure and manage.

The right unit is finished work: a task the AI completed that the next person in the process accepted without redoing it. A claims summary the adjuster uses as written is finished. One the adjuster rewrites is not, however many tokens it consumed. A code change that passes review and ships is finished. One that gets reverted is not.

That definition does two things. It forces quality into the measure, because rework can no longer hide. And it keeps the argument honest, because finished work can be counted. Shorter cycle times, fewer errors, and new revenue matter, but they are what finished work adds up to. You manage the unit and report the outcome.

The scoreboard is finished work per dollar, where the dollar includes the tokens, the tooling, and the people who check and correct the output. Cheap tokens on the wrong task are still waste. Some work needs a frontier model. Most does not.

Measuring it takes three things most AI programs never built.

1. A baseline

If nobody measured what a workflow cost before AI touched it, nobody can defend the investment afterward. The baseline does not need to be perfect. It needs to exist, and it needs to be agreed before the numbers come in, including what counts as finished for each workflow.

2. A map

Which models, agents, and copilots are running, on which jobs, at what unit cost, and who launched them. Most organizations find during this step that they are paying for work nobody remembers approving.

3. An owner

Someone with the authority to turn a model off, change a route, or stop a job that is not paying for itself. This is where most programs break down. As Arthur Riel has written about IT budgets, a manager who leaves spend alone faces no questions, while a manager who changes it takes the risk. AI spend follows the same incentives. Without a named owner and leadership backing to act, everything keeps running.

When I built a Fortune 500 insurance company's technology hub in Romania, we grew it to 600 people by holding every team to a cost baseline and a named owner. AI spend needs the same treatment. It is now a fast-growing line in the technology budget, and often the least governed.

The Levers Are Routing, Workflow, and People

The price of a token is not the lever. How you use it is. After its budget overrun, Uber says it brought its cost per token down while expanding access, by improving prompt caching, changing its default model, testing newer models for efficiency, and showing engineers their own usage and cost. None of that required a new vendor.

Routing

Many organizations send every request to the most expensive model they have. Classification, summarization, retrieval, and a large share of code assistance do not need it. Matching the model to the task, caching repeated context, batching work that does not need an immediate answer, and reserving the frontier model for hard cases changes the unit economics without changing vendors.

Workflow

Automating a weak process gives you a faster weak process. Tokens go up and finished work does not. The claims example makes the point: the largest saving came from redesigning the steps around the model so more output arrived finished, not from the model itself.

People

Applied AI engineering, security, data permissions, and operations redesign are scarce skills, and they are not getting cheaper the way general knowledge work is. Organizations that treat AI as a headcount cut usually stall. Organizations that treat it as an operating model problem still need people who can design the route, lock down the data, and read the scoreboard. That is where the investment belongs.

How We Assess an AI Estate in Five Days

We built the Coefficient Token Spend Monitor to make this work concrete.

It ingests OpenAI and Anthropic usage exports, plus invoice data where available, and puts them in one view: spend and tokens by provider, model, team, and workflow, with input, output, and cached usage separated. It reconciles against billing, flags missing data, and does not turn an instrumentation gap into an invented result.

Token data cannot tell you whether work was finished, so the assessment pairs usage with the measures leadership cares about: cost per finished task, quality, response time, and human rework. Where provider exports do not contain those measures, we record them with the client rather than assume them.

The assessment runs over five business days and covers up to three priority workflows. Day one sets the baseline and agrees on what finished means for each workflow. Days two through four analyze usage, connect spend to finished work, and rank the opportunities. Day five is a leadership readout with six outputs:

1. An AI spend baseline by provider, model, team, and workflow, including billing reconciliation and data gaps.
2. A prioritized opportunity register, with estimated savings ranges only where the data supports them, plus effort, dependencies, assumptions, and confidence.
3. Workflow recommendations covering model choice, prompts, context, caching, batching, retries, and validation.
4. A management scorecard for budget variance, cost per finished task, quality, response time, and rework.
5. A spending controls playbook with budgets, thresholds, accountable owners, escalation paths, and the action required at each threshold.
6. An executive readout and 90-day roadmap with owners, sequencing, and measures of success.

The monitor supports the assessment. It does not replace judgment. A rule can show that a workflow sends routine work to an expensive model, repeats uncached context, or loops on retries. It cannot decide whether a cheaper route still produces finished work. That takes a validation plan and an accountable owner.

We do not quote a savings number before we see the data. Each opportunity comes with the evidence, the method, and our confidence in it. Where the data is incomplete, we say so. Our AI Spend and ROI brief lays out the full sequence: assess the estate, run an optimization sprint, then put ownership and monitoring in place so spend does not drift back to volume for its own sake.

The Bottom Line

The companies that come out ahead in the next cycle will not be the ones that bought the most tokens. They will be the ones that measured what those tokens finished and rebuilt their operating model around the systems and people that produce it.

Before your next AI budget review, ask three questions. What did we spend, by workflow, and who owns each line? What share of the output was accepted without rework? What does a finished task cost now, compared with before AI touched it?

If your team cannot answer all three, the spend is not managed yet. If your AI spend is up and your operating results are flat, that is the place to start. Contact us.