What does a good answer cost?

The obvious AI metrics are cost per token and cost per call. They’re easy to get from the bill, and they rank models neatly: the small one is ten times cheaper. They also measure the wrong thing.

Nobody wants calls. A support team wants resolved tickets. An engineering team wants accepted changes. When the AI can’t finish, the work doesn’t disappear: a person picks it up, and that person costs far more than any model. So the unit that matters is cost per successful outcome, with the human follow-up included.

Measured that way, rankings flip. A model that resolves more tickets can be the cheapest option per ticket, even at ten times the price per call. This is unit economics again, with a harder question about what counts as one unit.

Compare the three models on cost per conversation, then turn on the human follow-up. Which model is cheapest now?

“Successful” needs a definition before it can be counted. In practice it’s a quality bar: an evaluation score, a human rating, a test that passes, a customer who doesn’t reopen the ticket. Outputs above the bar count as outcomes. The rest are cost without value.

Where the bar sits changes the economics. With a low bar, nearly every output passes and the cheap model wins easily. Raise it, and the cheap model’s pass rate collapses faster than the strong model’s. The cost of an outcome is the cost of a call divided by the pass rate, so a falling pass rate makes every success dearer.

That’s why model choice belongs to the people who own the outcome as much as to the people who own the bill. Set the bar first, then pick the cheapest model that clears it.

Slide the quality bar up from 50 and watch both pass rates. At what bar does the large model become cheaper per success?

A cost per outcome only means something next to a value per outcome. For most internal AI, that value is time saved: minutes per task, times tasks done, times what an hour of that person’s time costs. For customer-facing AI it may be revenue, retention or tickets avoided.

The full cost is more than tokens: licences, the engineering to build and run it, and the training and change work that gets people to use it. Adoption is the multiplier most estimates get wrong. A tool that saves thirty minutes a task but is used by a fifth of the team returns a fifth of the plan.

Time saved is also a soft number: it becomes value only if the hours go to other useful work. Treat the ROI as a range, and check it against something harder, like throughput, cycle time or tickets per agent.

Set adoption to 20%, then to 80%. Then find the minutes saved per task where the tool just pays for itself.

AI projects follow a J-curve. Months of building and piloting cost money before anyone gets value. Adoption then ramps slowly as people learn to trust the tool, and only after that does the cumulative line turn up. Time to value is how long you stay underwater.

Two things shorten it more than anything else. A small pilot that ships in weeks rather than a platform built for a year. And adoption work, which moves the ramp more than any model upgrade.

Time to value also decides how patient the budget has to be. A project that pays back in four months can be judged on results. One that needs eighteen has to be judged on leading indicators, such as weekly active users and outcomes per user, or it gets cancelled at the bottom of the curve.

Compare the big build with the pilot first. Then speed up adoption and see how many months it saves.

Most organisations run dozens of AI projects, each with its own team and its own success story. Put them on one chart, spend against value per dollar, and the stories sort themselves out.

Some initiatives create several dollars of value per dollar spent: scale them. Some are valuable but expensive: that’s where cost optimisation pays. Some cost more than they return and show no sign of improving. Stopping those funds the ones that work.

This is where FinOps for AI ends up, and the target is bigger than a cheaper token: value per dollar across everything the organisation spends on AI. Cost is the input. Value is the point.

Optimise the expensive initiatives, then stop the ones below the line. Watch value per dollar for the whole portfolio.