Per task
n/a
INSTRUMENT | Propagation & containment
Expected Cost of Failure Calculator for AI Agents
1 cited source
Price a failure rate in money: what gets caught, what escapes, and what correction and rework savings are worth, split by failure class.
A failure rate is not a cost until you price what gets caught, what escapes, and what correction takes. Grounded in what a silent failure costs when it looks like a success and how errors compound down a chain.
Savings mode reads the same arithmetic backward: cut a class's rate and see what it is worth, next to what one acceptable result costs and the bar your reliability needs to clear.
Showing your last valid result. Fix the error above to recompute.
- Occurrences
- 9,000
- Caught before delivery
- 5,850
- Escaped
- 3,150
- Detected share
- 65.0%
- Exposure hours
- 0 hrs
Enter a labor rate and a delay multiplier to price these occurrences. A delay multiplier of 1 means no penalty for a late fix.
Expected loss per month
Not priced yet: see the note above.
Escaped share of loss
n/a
Per failure class
| Failure class | Occurrences | Caught | Escaped | Caught cost | Escaped cost | Class loss | Share of loss |
|---|---|---|---|---|---|---|---|
| Wrong data extracted | 6,000 | 3,000 | 3,000 | n/a | n/a | n/a | n/a |
| Skipped a required check | 3,000 | 2,850 | 150 | n/a | n/a | n/a | n/a |
| (unlabeled) | 0 | 0 | 0 | n/a | n/a | n/a | n/a |
Savings mode read on the same numbers
| Savings mode | Value |
|---|---|
| Savings | n/a |
| Loss after savings | n/a |
| Net after change cost | n/a |
| Break-even change-cost share | n/a |
classLoss = caught x caughtCost + escaped x (M x correctionCost + downstream + remediation)How?
How this is calculated
Every failure class splits into what gets caught before delivery and what escapes. A caught occurrence costs correction labor, a reversal, and any auto-fix attempt. An escaped occurrence costs the same labor and reversal, but the delay multiplier scales that internal correction cost, plus whatever downstream impact and customer remediation the reader already priced for the escaped case. The multiplier never touches downstream or remediation a second time, since those figures already price the late, escaped outcome directly.
The per-caught and per-escaped figures above are per-occurrence dollars, never a period total. The per-escaped figure is the cost of one failure that reaches the customer after every check you run; paste it into the acceptable error rate calculator as its consequence cost, and the per-caught figure as its rework cost. Whether running that check is worth it in the first place is what the review threshold optimizer weighs, once both figures are in hand.
The delay multiplier. Boehm and Basili's measurement of defect-cost growth puts the ratio at about 100 to 1 on large projects and closer to 5 to 1 on small ones. This page asks for that ratio as a field rather than assuming one, because no single number fits every team's pipeline.
Detection hours. The hours to notice an escaped failure feed only an exposure figure, in hours. No money is priced from detection time here; that is the reader's own correction-labor field, priced once the failure is found. The same hours feed the error budget calculator's burn-rate window.
Savings mode. Cutting a class's failure rate by a stated percentage removes that share of its loss. The model does not change between modes, only the reading: savings is the same arithmetic looked at backward, against what a change to fix it would cost. Once you know what a failure costs, the acceptable error rate calculator inverts these same figures into the reliability bar automating would need to clear.
Formula: classLoss = caught x caughtCost + escaped x (M x correctionCost + downstream + remediation)
Two worked examples
A flat rate with no detection. 30,000 tasks a month, one failure class at a 30 percent share, no detection, a delay multiplier of 1, and a flat $10 remediation cost with everything else zero. That is 9,000 occurrences a month, all escaped, at $90,000.00 total and $3.00 per task. The figures are illustrative and reproduce no source this page names.
One model read twice. 12,000 tasks a month, one class at a 5 percent share, no detection, and a $20 remediation cost gives 600 occurrences and a $12,000.00 loss. Cutting that class's rate by 40 percent removes 40 percent of its loss: $4,800.00 in savings, leaving $7,200.00. The numbers are this page's own illustration.