ANSWER

Who is accountable when an AI agent makes a mistake?

Reviewed August 1, 2026.

The company operating the process is accountable — before customers, auditors and data-protection law, saying the AI did it transfers nothing. In personal-data processing, controller and operator answer for what the agent does. The useful question is not whose fault it is, but whether the design can answer: every action recorded, every exception owned, every irreversible step approved by a named person, and an immediate shutdown always available.

Step by step

• Name the human owner of each agent before the first run — a role and a person, not a department.
• Guarantee a per-action trail: who, what, when and under which rule. Without it, every error becomes archaeology.
• Put human approval on irreversible steps: payments, deletion, production changes, external communication.
• Treat exceptions as a queue with an owner and a deadline, not an alert scrolling by in a channel.
• Rehearse the incident before you need it: who shuts down, who communicates, who decides the reversal.

How an agent does this

This is why AgenticosCore does not operate an agent without the four controls: the Trail (per-action record), the Brake (one-click shutdown), House Rules (who approves what) and the Parallel run (proof before cutover). The agent executes what is reversible; the irreversible decision always belongs to a named person.