Reversal paths and audit trails

The question a board asks after an AI incident is not "why did the model do that?" It is "can we undo it, and can we see what happened?" If both answers are yes, the incident is a Tuesday. If either is no, it is a crisis.

Reversal paths

• Snapshot before acting. The price before the change. The ticket state before the reply. The ledger before the refund.
• A rollback window. 24 hours is a good default. Within it, one click restores the snapshot.
• Batch rollback. If an agent made 400 bad decisions overnight, you need to reverse them as a group, by agent and time range, not one at a time.

The audit record

For every decision, store: the inputs the agent saw, the proposal it made, every guardrail that ran and its result, the final action, who or what approved it, and the snapshot. Keep it in a place the agent cannot write to.

{
  "decision_id": "px-20261017-0412",
  "agent": "pricing",
  "inputs": {"demand_index": 0.84, "competitor_gap": -4},
  "proposed": {"change_pct": 11},
  "guardrails": [{"step_limit": "clamped to 8"}, {"floor": "ok"}, {"parity": "ok"}],
  "action": {"change_pct": 8, "new_price": 412},
  "snapshot": {"old_price": 381},
  "reversible_until": "2026-10-18T04:12:00+04:00"
}

This record is also what makes governance possible. Compliance, finance and customers can all be answered from it.