Engineering, production, governance, cost, communication · insight-0277
Make every automated decision reconstructable from a decision record
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- ML engineer, Data scientist
- Industries
- banking, insurance, healthcare
Summary: situation, action, reason, boundary
Situation
A customer disputes a declined credit-limit increase from eight months ago, and the team must explain it.
Action
For each decision, store a request id, timestamp, model and rule versions, the feature values or a pointer to their point-in-time snapshot, the score, threshold and final action, with a retention period agreed with legal.
Reason
Auditability requires recreating what the system knew and did at decision time, not what it would do now.
Boundary
Credit, insurance, pricing, hiring, healthcare triage and other consequential decisions.
Symptom or risk
Without the inputs and version used at the time, the team re-scores with today's model and data and gives an answer that may be wrong.
Exceptions
Low-stakes personalisation may keep only aggregated logs.
Trade-offs
Decision logs contain personal data, so they need access controls and retention limits.
Pitfalls
Logging the score but not the threshold or business rules applied after it.
Lessons and concepts
Claims and their sources
The NIST AI RMF organises AI risk management into Govern, Map, Measure and Manage functions and lists accountability and transparency among trustworthy-AI characteristics.