Classify failures into three types before analysis — preventable, complex, or intelligent

Why This Is a Rule

Amy Edmondson's failure taxonomy distinguishes three types that require fundamentally different analytical questions:

Preventable failures (process deviation): a known process existed and wasn't followed. The question is "why wasn't the process followed?" not "what went wrong?" The fix is process compliance, not process redesign. Example: deploying without running the test suite despite a documented deploy checklist.

Complex failures (novel factor interaction): multiple factors combined in ways that couldn't have been predicted from any individual factor. The question is "what novel interaction produced this?" not "who was responsible?" The fix is better detection and system design, not blame. Example: a database migration interacting with a cache invalidation bug that only manifests under specific traffic patterns.

Intelligent failures (frontier experiments): the failure occurred at the boundary of current knowledge, where failure was a possible and acceptable outcome. The question is "what did we learn?" not "how do we prevent this?" The fix is to extract the learning and feed it forward. Example: A/B test that showed the new feature decreased engagement.

Applying the wrong analysis to the wrong type wastes effort and misattributes causes. Analyzing a complex failure as if it were preventable produces blame for something no one could have foreseen. Analyzing an intelligent failure as if it were preventable produces risk aversion that kills experimentation.

When This Fires

• At the start of any failure analysis, post-mortem, or retrospective
• Before asking "what went wrong?" — the type determines the right question
• When a failure has occurred and the team is deciding how to respond
• During any learning-from-failure practice

Common Failure Mode

Treating all failures as preventable: "Someone should have caught this." This is only correct for Type 1 (process deviation). For Type 2 (complex), no one could have predicted the interaction. For Type 3 (intelligent), the failure was the expected possible outcome of a deliberate experiment. Defaulting to preventable analysis produces blame culture and kills both innovation and honest reporting.

The Protocol

At the start of every failure analysis: (1) Classify: was this a preventable failure (known process not followed), a complex failure (novel factor interaction), or an intelligent failure (frontier experiment)? (2) Ask the type-appropriate question: Preventable → "Why wasn't the process followed?" Complex → "What novel interaction occurred?" Intelligent → "What did we learn?" (3) Apply the type-appropriate response: Preventable → improve process compliance. Complex → improve detection and design. Intelligent → extract learning and continue experimenting.