Back to library

AI Explainability Requirements

Explainability is the product's answer to "why did the AI do that?" Users who understand why an AI made a recommendation are more likely to trust it, catch its errors, and use it effectively. Regulators and enterprise buyers increasingly require it. This skill defines the explainability requirements for a feature and the product design that delivers them.

---

Context

The three levels of explainability:
LevelWhat it providesWho needs it
Output-level"Here's what the AI produced and a confidence indicator"All users
Decision-level"Here's why the AI made this recommendation"Users affected by AI decisions; regulated contexts
Audit-level"Here's a full record of the AI's inputs, reasoning, and outputs"Enterprise buyers; compliance teams
The explainability tension:

The most accurate AI models are often the least interpretable. The PM must decide if the accuracy benefit is worth the explainability cost.

---

Step 1 — Define the explainability requirements

Ask: what is the AI deciding, who is affected, regulatory context, what users need to use the output responsibly, what they need to challenge it, what auditors need.

Step 2 — Select the explainability level

Output-level (minimum for all AI features), Decision-level (High/Critical stakes), or Audit-level (regulated/enterprise contexts).

Step 3 — Design the explanation content

For recommendations: top 3 factors in plain language + counter-factual.

For classifications: result + primary evidence + what would change it.

For generative features: source citations + confidence indicators.

Step 4 — Design the explanation UX

Patterns: simple reason tag, ranked factors, confidence + source, or decision audit panel. Plus challenge/override UI for all decision-level features.

Step 5 — Define the audit log requirements

For audit-level: full log schema including model version, input hash, output, explanation, human reviewer actions, user challenges. Retention and access controls.

Quality check before delivering

Explainability level is appropriate to the stakes
Explanation content uses plain language examples
Counter-factual explanation is included for decision-level features
Challenge and override pathway is specified
Regulatory context is addressed
Non-technical user readability is a checklist item
Suggested next step: Test the explanation with someone who wasn't involved in building the feature. Ask: "Does this tell you why the AI did that? Would you know what to change if you disagreed?" If they can't answer both, the explanation isn't working.