How to Budget AI Agents by Cost per Finished Task, Before the Bill Arrives

Budget an AI agent by what it costs to complete one task to an agreed standard, and use the pilot to measure that number and to test what happens when spending reaches its limit. More AI software is now billed by usage, and cost is already a common reason projects stall. In a small VentureBeat survey published on November 10, 2026, 38% of respondents said their organization had delayed, scaled back or canceled an AI project because it cost more than expected, against 12% who cited insufficient computing capacity.

In the preceding two weeks, Oracle announced usage charges on top of its subscriptions, GitHub said Copilot for enterprises would be billed through AI credits, and Microsoft began limiting free Copilot Chat usage. Each change ties what a customer pays, or how much a user can do, to the amount of AI work done.

Who charges for what

A seat license is a fixed price per named user per month, and it is easy to forecast because headcount is. Consumption pricing charges for activity, and an agent's activity depends on how many steps it takes to finish a job.

For a comparison of how agent vendors bill per conversation, per action and per resolved case, see AI agent pricing models.

What the 38% figure shows, and what it does not

The number comes from VentureBeat's VB Pulse survey on AI infrastructure, fielded in October 2026 with 156 qualified respondents, described as people who work on AI at, or consult for, organizations with at least 100 employees. It shows that these respondents cited cost overruns about three times as often as capacity shortages, and that only 26% said their organization judges AI spending by a measurable financial return. Thirty-eight percent of 156 is roughly 59 people. The article does not describe how respondents were recruited or give the exact question wording. The survey covers AI projects in general, does not single out agents, and does not establish whether usage charges caused the overruns. Read it as a signal that cost belongs in the pilot plan.

The unit to budget: cost per finished task

A finished task is one completed to a written acceptance standard, including any human approval the process is designed to have. A task that a person has to redo does not count. The full cost per finished task is everything spent on the agent divided by the number of finished tasks, and "everything" covers usage charges, licenses, build, hosting, monitoring and review time. Finance can compare that figure with what the task costs today, which raw usage totals do not allow. An Amazon engineering memo published on October 27, 2026 called AI usage metrics "deeply flawed" because "they are disconnected from what matters".

Usage is the part of the cost that vendors now meter and the part that moves with volume, so estimate it first. Use these four inputs to build an initial range, then measure them on representative tasks.

1. Steps per task. Count every call the agent makes to a model. Microsoft's billing documentation notes that to answer one message an agent can call the model several times, and "each call counts".
2. Billed tokens per step. A token is a fragment of a word, and models are metered by tokens read and written. Each call re-sends the agent's instructions, its tool descriptions, the conversation so far and whatever it has retrieved.
3. The model at each step. Rates differ by model, so the estimate needs a model named against every step.
4. Retries and acceptance rate. Some tasks need a second attempt, and some are still rejected or redone by a person.

Worked example (illustrative task and volumes, at the Copilot Studio rates in Microsoft's billing documentation). An agent answering a supplier query makes four calls to a reasoning model at 5,000 billed tokens each, which is 200 credits, takes three agent actions at 5 credits each, and returns one generative answer at 2 credits. One attempt uses 217 credits. If one task in five needs a second full attempt, the average is about 260. If 85% of tasks are accepted, usage per finished task is about 306 credits, 41% above the single-attempt figure. Licenses, build cost and the staff time spent on the rejected tasks are extra.

During the pilot, replace each assumption with a measurement. Report the typical task and the most expensive tenth separately, because a few long tasks can account for a large share of the spend.

Four factors the sources say move the cost

None of the sources ranks these factors against one another, and most of the figures come from vendors, so a pilot should measure all four on its own workload.

The model used on each step. A Gartner note, as reported by The Register on November 3, 2026, says nine in ten agent workflow steps are routine and that open-weight models can run them for as little as a twentieth of the cost of a top-tier model. Gartner's own estimate of the realistic saving is five to six times per task, because those models "can be more verbose and retry more often".

The effort or depth setting. Anthropic's own figures for Claude Fable 5.2, as reported by VentureBeat on November 9, 2026, show a coding benchmark task costing nearly six times as much at the Xhigh effort setting as at Medium, while the score rose from 37.6% to 49.1%. GitHub's estimate for its deepest Copilot code review is about 40 times its estimate for the lightest, according to InfoWorld on October 30.

How much context each call carries. In Microsoft's illustrative example, the token charge for the same follow-up question is 141 credits inside a running conversation and 461 credits when it opens a new one, because less of the prompt can be read from the cache. Feature charges are added to both figures.

Retries and the harness. The harness is the software around the model that runs the loop of steps. AWS's Strands team reported in October that across about 94,000 benchmark trials, token usage varied by roughly two times depending on the harness, often with little difference in accuracy. The team builds one of the harnesses tested and estimated its dollar costs from token counts. The Register has also noted developer complaints about runaway costs in AWS's Kiro coding agent, which bills in credits whose token value depends on the model.

What the caps do, and what to ask before signing

First establish whether a vendor control stops spending or only reports it. On GitHub, a spending limit sends a notification and charges continue to accrue unless "Stop usage when budget limit is reached" is enabled, while a budget set on an individual user always stops that user. Google says a hard limit can be set on a project's AI spend and the project's agent pauses when it is reached. The post does not state its availability.

Microsoft Copilot Studio blocks an environment's agents when a tenant reaches 125% of its prepaid capacity, and a blocked agent stops responding even in the middle of a conversation. With a pay-as-you-go plan the capacity block does not apply and the excess is billed, although an agent that has reached a monthly stop limit set by an admin stays blocked. That monthly limit does not stop the agent flows the agent calls. Meters and layered limits are covered in the guide to agent cost monitoring.

Ask each shortlisted vendor to answer these six questions in writing before a contract is signed.

• What is the billing unit, and how many units does one finished task of ours consume when measured on our own data?
• Which controls stop spending and which only alert, at what level, and what happens to work already in progress?
• Who can change the default model or effort setting, and how much notice do we get before it changes?
• Are failed runs, retries and test runs billed, and at what rate compared with successful runs?
• Can unit rates, included allowances or usage limits change during the contract term?
• Can we export usage by agent, by user and by model, in a form finance can reconcile to the invoice?

Seats can remain the better deal where each person's use is heavy and steady. Microsoft's rate table shows no credit charge for most agent use by a signed-in Microsoft 365 Copilot license holder, though agents that run autonomously still consume credits. The advice here would also change if a vendor offered a per-result price with a definition of "result" the buyer can audit, because the vendor then carries the variance and the token estimate matters less. An automatic stop is the wrong default for an agent whose outage costs more than its overrun. That agent needs an alert threshold, a named owner, a bounded emergency allowance and a manual fallback.

What to do this month

1. List every AI product you already pay for and write down its billing unit and whether paid overage is switched on by default.
2. For the one agent closest to production, estimate usage per finished task, add the fixed costs and compare the total with the current cost of that task.
3. Turn on a hard stop for every pilot and test environment, and write a fallback plan for each production agent that is exempt from automatic stops.
4. Agree with finance that agent business cases will be reviewed on cost per finished task, with the acceptance standard and measurement method written into the pilot plan.

Cost-per-task measurement can be scoped into QueryNow's AI agent work alongside the accuracy testing a pilot already needs.