Agent Audit Checklist

Use this as a pre-flight list. The reasoning behind each item is in the permission model and the confidence thresholds playbook.

Scope

• Is there a written list of fields the agent may write, and a block list of fields it may not?
• Does the block list include certifications, compliance declarations, hazard data and country of origin?
• Does the agent write through the system’s own interface, or by importing files?

Evidence

• Does every proposed value carry a source document or system?
• Is each field’s evidence state known: Known, Missing, Conflicting, Unproven or Stale? See evidence gap.
• Is identity (which product or supplier) confirmed before any attribute is written?

Permissions

• Is each field tier mapped to one of: execute, write with flag, recommend, request evidence, block?
• Are thresholds calibrated against sampled human review, not set from the model’s own score?
• What happens when sources conflict?

Audit trail and reversal

• Is each write-back logged with old value, new value, source, confidence, rule and approver?
• Can a write, or a whole batch, be reversed?
• Is there a sampled review of automatic writes, and a way to pause the agent?

FAQ

What should you check before an AI agent writes to product records? +

Which fields it may write, what evidence each write needs, what confidence and conflict rules apply, how every write is logged, and how a bad write is reversed.

Which fields should an agent never write on its own? +

Fields with legal, safety or contractual weight: certifications, compliance declarations, hazard data and country of origin. These need an authoritative source or human approval.