REFERENCE · §4 · LAST REVIEWED 2026-04-27

ACF §4 — Reversibility

Reversibility is the technical and procedural capacity to undo, compensate for, or contain the effects of an autonomous agent action — within a defined window, by a defined operator, against a defined runbook — without depending on cooperation from the agent itself.

A regulator asking "what happens when your agent makes a mistake" is asking about reversibility. The honest answer for most firms today: nothing structured. The framework section requires a documented window (e.g., 30 days for trade reversal, 24 hours for KYC re-screen), an authorised operator, and a runbook tested at least quarterly. Maps to MiFID II best-execution recovery, FCA Principle 6 (treating customers fairly), and DORA operational-resilience drills.

Regulatory anchors

• MiFID II Best Execution
• FCA PRIN 6
• DORA Art. 24-25
• EU AI Act Art. 14(4)
• GDPR Art. 22(3)

What this covers

• Reversal windows per action class — trade reversal, payment recall, KYC redo
• Authorised operator and dual-control where applicable
• Compensating-control mechanisms when full reversal is impossible
• Customer-facing communication for reversed actions
• Quarterly reversal-drill testing

Common gaps

• No explicit window per action class — informal "we’ll fix it" culture
• Reversal possible in code but never tested under realistic load
• Compensating controls undocumented
• Customer comms scripts do not exist for "the agent made a mistake" scenario

Related sections

• §2 — Action Perimeter

  What an agent must NOT do — technical and policy guardrails on action space.

• §3 — Audit Trail

  Every agent action logged with sufficient detail to reconstruct intent and outcome.

• §9 — Customer Disclosure

  When and how to disclose to end-users that an agent is taking actions on their behalf.

• §12 — Incident Response

  When the agent does the wrong thing — runbook, regulator notification, customer remediation.

Reference compiled by Sebastian Heine. Editorial perspective at The SHeine Brief.