AI Agent Costs: Measure Per Outcome, Not Per Token

Uber published the economics of running AI agents across its software organization this week, and the headline number is worth reading twice. Between February and July 2026, agent requests grew roughly 9.4x, yet cost per session fell 52 percent and cost per 1,000 model requests fell almost 34 percent. Usage went up nearly tenfold and the unit cost of the work went down by half. The interesting part is not the model choices. It is how Uber decided what to measure: cost per merged pull request, per code review, per triaged alert, paired with quality signals like revert rate. Cost denominated in outcomes, not tokens. Most businesses running agents are still watching the wrong number.

The Problem: A Bill That Tells You Nothing

If you run an AI agent today, the number you probably see is a monthly invoice from your model provider, or a token dashboard. It goes up when usage goes up. That is all it can tell you. It cannot say whether the agent resolved 900 support tickets or 90, whether a cheaper model would have done the same job, or whether the spend is a bargain or a leak. A rising bill can mean a healthy, growing workload or a runaway loop, and the dashboard shows both as the same line.

This is why so many teams end up in one of two bad places. Either they cut spend blindly, downgrading models and throttling usage until quality drops and the agent stops being useful, or they let it climb because nobody can prove it is too high. Both come from the same gap: no measure that connects money spent to work finished.

What Cost Per Outcome Actually Means

Cost per outcome is the total spend to produce one unit of completed, accepted work. Not one API call, one conversation, or one thousand tokens, but one thing your business actually wanted done. The unit depends on the agent:

• An invoice-processing agent: cost per invoice posted correctly without human correction
• A voice agent: cost per booked appointment or per resolved call, not per minute of audio
• A support agent: cost per ticket closed that did not reopen within a week
• A lead-qualification agent: cost per qualified lead handed to sales
• An internal ops agent: cost per completed request that a human did not have to redo

The second half of each definition matters as much as the first. Uber pairs cost with revert rate and review accuracy for a reason. An agent that is cheap per attempt but wrong a third of the time is not cheap, because every failure is paid for twice: once in tokens and again in the human time spent fixing it. Counting only accepted outcomes puts that hidden cost back into the number.

Why This Metric Unlocks Real Savings

Once cost is tied to outcomes, the levers become visible and testable instead of guesswork. Uber benchmarks models against real internal tasks and routes each kind of work to whichever option gives the best mix of quality and cost, discarding configurations that are worse on both. You can do the same at a much smaller scale. Take a sample of 100 real tasks, run them through two or three model setups, and compare cost per accepted outcome. Often a smaller model handles the routine 70 percent of the work at a fraction of the price, and a stronger one is reserved for the hard cases. Without the outcome metric, that comparison is impossible, because per-token price alone will always favor the cheapest model even when it fails more.

It also exposes waste that no token dashboard can. Agents that retry the same failing step, pull far more context than the task needs, or loop on ambiguous inputs all show up as a rising cost per outcome long before they show up as a shocking invoice. That matters because monitoring is thin in most companies. New Relic data published this week found that roughly one in four AI agents runs unmonitored. An agent nobody is watching cannot be optimized, and its cost per outcome is an unknown.

A Realistic Scenario

Consider a regional logistics broker with about 80 staff that deployed an AI agent to read carrier invoices and match them against shipment records. After three months the model bill was around 4,200 dollars a month and management assumed that was simply the price of the system. Nobody could say whether it was good value. Wizeb added two measurements: how many invoices were posted with no human correction, and how many minutes staff still spent fixing the rest. The picture changed immediately. Only 71 percent of invoices posted cleanly, and the corrections were eating about 60 staff hours a month. Counting review time, the true cost per correctly posted invoice was far higher than the token bill implied.

The fixes were unglamorous. Wizeb routed clean, standard invoices to a smaller model, sent only ambiguous ones to a stronger model, cached the carrier rate tables the agent kept re-reading, and added a validation step that caught mismatched totals before posting. Clean-post rate rose above 90 percent, correction time fell by more than half, and the monthly model bill dropped as well. The company now tracks one number weekly, cost per correctly posted invoice, and can show it to the finance team.

How to Start Measuring This Week

1. 1Define the outcome. For each agent, write down the single unit of finished work it exists to produce, and what counts as accepted without rework.
2. 2Log outcomes, not just calls. Record each task with a result: accepted, corrected by a human, or failed. Without this, the denominator does not exist.
3. 3Add human time to the numerator. Even a rough estimate of minutes spent reviewing or fixing is far better than ignoring it.
4. 4Set a baseline and a same-week alert. Track cost per outcome on a trailing seven-day average so drift shows up in days, not at month end.
5. 5Test one lever at a time. Try a cheaper model on routine cases, trim context, or add a validation step, and keep the change only if cost per outcome improves without quality falling.

How Wizeb Approaches This

Every agent Wizeb builds ships with outcome logging from day one, because an agent you cannot measure is an agent you cannot improve or defend to a finance team. We define the unit of work with you, instrument accepted versus corrected results, set a cost-per-outcome baseline, and then use it to choose the right model for each class of task and to catch cost drift early. For agents already in production, we run a cost-per-outcome review that usually finds savings in model routing, context size, and retry behavior without touching quality. If you are running agents and the only cost figure you have is a provider invoice, that is the place to start. See how at wizeb.com/services/ai-agents.

Three Questions to Ask About Every Agent You Run

1. 1Can you state, in one sentence, what one finished unit of work is for this agent, and how many it completed last week?
2. 2Do you know what share of its output a human had to correct or redo, and what that time cost?
3. 3If you swapped in a cheaper model for the routine cases, would you be able to tell within a week whether it was a real saving?