Definition
Blast-radius exposure created by AI-selected deployment strategies, weighing speed gains against the user-minutes of exposure incurred when a fast strategy meets a bad release.
The Measurement Gap
Measurement Gap and Decision
The Gap: Traditional deployment metrics celebrate speed gains without accounting for exposure costs. "Average deployment time reduced by 40%!" doesn't mention that the three failures this quarter affected 50x more users than under conservative defaults. Mean time to recovery looks good because rollback was fast, but users already experienced the failure during the exposure window the conservative strategy would have closed.
The Decision:
Is the AI's deployment strategy optimization (faster canary, shorter monitoring windows, aggressive rollout) producing net positive or net negative user outcomes compared to a conservative default? A positive DSEI means the AI's optimization exposed more users to failures. A negative DSEI means it actually protected more users by detecting failures faster.
What Changes When You Have This Number:
Below 50 user-minutes per release average, the velocity gains likely justify marginal exposure increase. Between 50–200, investigate which release categories drive the cost. Above 200, cap AI-recommended rollout speeds: the speed gains don't justify the exposure. You can also compute the explicit trade-off: are 75 engineer-hours of deployment time saved worth 2,500 user-minutes of failure exposure?
Formula
Formula:
DSEI = Σ (Users_affected_AI_strategy − Users_affected_conservative) / N_releases
Where Users_affected_AI_strategy is users who experienced errors during failed releases under AI-recommended canary/rollout parameters (measured in user-minutes), Users_affected_conservative is estimated users who would have experienced errors under conservative default strategy (counterfactual calculation), and N_releases is total releases in the period (successful releases contribute 0).
Sign Convention:
Positive means AI strategy exposed more users. Negative means AI strategy protected users better. Zero means equivalent exposure or no failures occurred.
Worked Example:
Quarter with 100 releases, 3 failures:
Failure 1: AI strategy (10% canary, 5 min → 100%) exposed 8,000 user-minutes. Conservative (1% canary, 30 min staged) would have caught at 1%, exposing 120 user-minutes. Delta: +7,880.
Failure 2: AI strategy exposed 3,200 user-minutes. Conservative: 800 user-minutes. Delta: +2,400.
Failure 3: AI strategy exposed 1,500 user-minutes. Conservative: 1,200 user-minutes. Delta: +300.
DSEI = (7,880 + 2,400 + 300) / 100 = 105.8 user-minutes per release
This falls in the moderate range (50–200). Failure 1 dominates: the AI used a very aggressive rollout that missed a defect only visible under production load patterns the 5-minute canary window didn't capture.
Conservative Baseline:
Must be defined explicitly and held constant. Typical default: 1% for 30 min → 10% for 30 min → 25% for 15 min → 50% for 15 min → 100%. Total ~90 minutes.
Proposed formula, pending validation pilots.