Decision Audit

Simple Definition

A decision audit is a review of the assumptions, interpretations, shortcuts, and tradeoffs an AI agent made while completing a task, conducted after the work is done, to surface choices that were never explicitly reviewed.

Decision Audit vs. Decision Mapping

These sit on opposite ends of the same process. Decision mapping happens before work starts, separating what’s already decided from what’s still unknown, so the agent doesn’t rush into building on unresolved questions. A decision audit happens after the work is done, reviewing the choices the agent actually made along the way, including ones nobody explicitly approved.

Why It Matters

An agent working through a multi-step task makes dozens of small decisions that never get surfaced to a human: which interpretation of an ambiguous instruction to use, which edge case to skip, which library or approach to default to. Most of these are fine. Some aren’t, and without an audit, they ship silently inside a result that otherwise looks complete.

Example

After a coding agent finishes a feature, a developer asks it to list every non-obvious decision it made: which error-handling approach it chose, what it assumed about an ambiguous requirement, and any edge case it deliberately didn’t handle. Reviewing that list surfaces two assumptions the developer disagrees with, caught before the code ships instead of after.

A Simple Decision-Audit Template

What did I assume that wasn’t explicitly stated?
What ambiguous instruction did I interpret, and how?
What edge case or requirement did I skip or simplify?
What would I have asked a human about, if I’d had the chance?
What’s the one decision here most likely to be wrong?

Decision Mapping, the equivalent process done before work starts
Verification Gate, checking the output itself, distinct from auditing the choices behind it
Pride Gate, a related reflective check on whether the agent stands behind its own work
Blind-Spot Audit, surfacing what the agent is unsure about