DemonstrationInteractive tool

Agent pilot cost and reliability model

A calculator for whether an agentic AI pilot is worth building, once you count the tokens a failed run still burns and the human review time it still requires.

What it shows

Pricing an AI pilot on completed outcomes, not attempted runs
Making retry cost and residual human review time visible
Finding the volume where a pilot actually breaks even

A demonstration built from public and synthetic data, not from any real client or organization.

The cost of a completed, acceptable outcome, and the human time the pilot still consumes after it ships, are the two numbers that decide whether an agent pilot is worth building. Both run higher than the per-run token estimate a pilot usually gets approved and renewed on.

This model asks for the inputs a team already knows: how many runs a month, how many steps a run, roughly how many tokens a step, what share of runs a person still reads, and how long that reading takes. It then separates cost per attempt from cost per success, because a run that fails still spends tokens and still spends reviewer minutes without producing anything.

The starting numbers are sourced: token prices match current published API pricing and the hourly rate matches the U.S. Bureau of Labor Statistics’ latest measure of loaded labor cost, both cited on the page. Enter your own numbers if your situation differs.

Run volume and size

Count every run you kick off, including retries. Failed runs still cost money, and the model needs to see them.

Attempted runs per monthRuns per month, counting every run you start including retries.

Model steps per runSteps per attempted run. One step is one call to the model.

Input tokens per stepAverage input tokens sent on a step, including context resent from earlier steps.

Output tokens per stepAverage output tokens produced on a step.

Model pricing

The starting values come from current published API pricing, cited in the sources below. Enter your own numbers if your provider or model differs.

Input priceDollars per million input tokens, as your provider bills it.

Output priceDollars per million output tokens, as your provider bills it.

Human review

Review is counted against attempted runs, not completed ones, because reading the output is how a bad run gets caught. The starting hourly rate comes from a published labor statistic, cited in the sources below.

Share of runs a person readsPercent of attempted runs that get human review, from 0 to 100.

Minutes per reviewMinutes a reviewer spends on one reviewed run.

Reviewer hourly costDollars per hour, loaded, for whoever does the review.

Failure ratePercent of attempted runs that do not produce an acceptable outcome, from 0 to 100.

The manual baseline

What the same work costs today, done by a person. Uses the same sourced hourly rate as human review by default.

Minutes per task by handMinutes a person takes to do one of these tasks manually, start to finish.

Hourly cost of that personDollars per hour, loaded, for whoever does the task today.

Build cost

What it takes to get the pilot running, spread over the period you expect it to last.

One time build costTotal dollars to design, build, and ship the pilot.

Spread the build overMonths. How long you expect this pilot to run before it is rebuilt or retired.

Your result

Worth piloting

At these numbers the pilot costs less per completed outcome than doing the work by hand, with enough room that a modest miss in your estimates would not flip the answer.

each completed outcome costs 6.90 dollars against 20.55 dollars to do by hand
you are running about 5.6 times the break even volume of 76 completed outcomes a month
review takes 33.3 hours a month against 177.1 hours by hand, freeing about 143.8 hours

Cost per completed outcome

$6.90

Against $20.55 by hand.

Cost per attempt

$5.87

What every run costs, whether or not it works.

Attempts per completed outcome

1.18

Attempts you pay for to get one result.

Monthly cost of the pilot compared with the manual baseline
Every month	Amount
Model spend	$39.00
Human review, 33.3 hours	$1,644.00
Build, spread over 12 months	$1,250.00
Total pilot cost	$2,933.00
Same work by hand, 177.1 hours	$8,733.75
Saved	$5,800.75
Human hours freed	143.8
Break even	About 76 completed outcomes a month

Want a second opinion on these numbers?

Send us what you entered and we’ll tell you which assumption we would test first. We only use your email to reply.

What is the workflow? (optional)

Where the starting numbers come from

Anthropic, Claude API pricing: Claude Sonnet 5: $2 / $10 per million input/output tokens, introductory pricing through August 31, 2026 (standard pricing is $3 / $15 after). Checked August 2026.
Google, Gemini API pricing: Gemini 2.5 Pro: $1.25 / $10.00 per million input/output tokens for prompts up to 200k tokens. Checked August 2026.
OpenAI, GPT-4o pricing: Widely reported at $2.50 / $10.00 per million input/output tokens. OpenAI does not publish a static pricing page we could cite directly; this figure is corroborated across multiple independent trackers. Checked August 2026.
U.S. Bureau of Labor Statistics, Employer Costs for Employee Compensation: Average total compensation (wages plus benefits) for all civilian workers: $49.32 per hour, March 2026. Used as the starting loaded hourly rate for both the reviewer and the manual baseline.

A demonstration. The starting numbers are cited (see sources above), but every field is yours to edit. This is a planning aid, not a forecast, and it does not provide financial advice.

← All demos

Bring us the decision, process, or system that is not working.

We will help you understand the problem, determine what the evidence supports, and build a better way forward.