Skip to content

Agentic operations

Agentic Workflow Breakeven Calculator

Model an agentic workflow against the task it replaces: tokens with retry overhead, the human audit sample, rework of failures and the cost of what ships wrong — then read the failure rate at which the saving reverses.

Runs in your browser Audited by Gini Loh
Instrument panel Deterministic · nothing leaves the page
The task it replaces3
runs

Runs started per month, including the ones that fail.

min
$/hr
Token economics3
tok

Prompt, completion and retrieved context, per run.

$/1M

Share of runs, as a decimal: 0.22 is 22%.

Human in the loop4

Contract or retrieval misses per run, as a decimal: 0.057 is 5.7%.

Share of runs a person reads, as a decimal: 0.25 is 25%.

min
min
Reviewer and exposure2
$/hr

The person who audits and reworks, not the operator.

$/run

SLA credit, refund or remediation for a failure nobody reviewed.

Cost per run$19.90+$6.40 vs manual
Verification cliff3.2%3.2% failure
Monthly delta$160kagainst manual handling

Above the verification cliff

At 5.7% failure the agent costs $19.9 against $13.5 manually — 47.4% more per run, $159,942 more a month.

The cliff sits at 3.2%. Every failure point above it adds $254 per run.

Cost per run against failure rate5.7% assumed
Manual handlingAgentic run
Cost per run by failure rate
Failure rateAgent costManual cost
0.0%$5.41$13.50
2.5%$11.76$13.50
5.0%$18.12$13.50
7.5%$24.47$13.50
10.0%$30.83$13.50
12.5%$37.18$13.50
15.0%$43.54$13.50
17.5%$49.89$13.50
20.0%$56.24$13.50
22.5%$62.60$13.50
25.0%$68.95$13.50
Where the money goesone run at the assumed failure rate
Tokens, including retries$0.08
Human audit sample$5.33
Rework of failures$3.80
Escaped defects$10.69
Manual handling$13.50
  • Rework is charged on every failure and escaped failures pay both rework and penalty. Audited failures are assumed to be caught, not shipped.

When the invoice is the question

This estimate takes your retry overhead on trust. SpendProof reconciles an AI provider invoice against your own tagged usage and withholds a clean close when a bucket does not tie out.

Open SpendProof ↗

Operator notes

How to run it

  1. 1. Start with the manual task: minutes, loaded cost and how many runs a month.
  2. 2. Enter the failure rate you actually measure, not the one in the vendor deck.
  3. 3. Set the audit sample to the share a person really reads today.
  4. 4. Price the reviewer at the rate of whoever audits and reworks, not the operator.
  5. 5. Read the cliff, then compare it with the failure rate on line two.

Model logic

Manual cost = minutes ÷ 60 × loaded rate. Agent cost = tokens × (1 + retry) × price, plus the audit sample, plus rework on every failure, plus the penalty on the share nobody reviewed. The cliff solves that expression for the failure rate where it equals the manual cost.

FAQ

What is the verification cliff?

The failure rate at which an agent costs the same per task as doing it by hand. Above it, every extra point of failure costs more than the labour the agent removed, so the project gets more expensive the harder it is pushed.

Why does the audit sample dominate the token line?

Tokens are a rounding error beside human minutes. Fourteen thousand tokens at $4.50 per million is under eight cents; eight minutes of a $160/hour reviewer on a quarter of the runs is over five dollars. A per-token comparison and a per-task comparison are different questions.

Is an escaped defect charged twice?

It pays its rework and its penalty. Someone eventually finds it and fixes it, so rework is charged against every failure rather than only the audited share, and the penalty is charged against the share nobody reviewed.

What does the engine refuse to assume?

That a cliff always exists. When tokens plus the audit sample already exceed the manual cost there is no failure rate at which the agent wins, and the page says so instead of printing a small margin.

Related dispatch

Reducing Retrieval Misses. A technical isometric cutaway of a tiered mechanical sorting funnel filtering wide blocks into a narrow brass tray.

Your Retrieval Fails at 5.7%

Anthropic's retrieval test missed the right passage 5.7% of the time when it fetched the 20 best-matching chunks.

Read dispatch