Who is accountable when the agent gets it wrong?

An assistant that suggests a reply leaves a person in charge. An agent that sends the email, books the meeting or changes the configuration does not, at least not in the moment. As more products move from suggestion to action, the question of who is responsible when something goes wrong has stopped being theoretical.

Elena Petrova chairs a panel of people who have had to answer it. Zara Haddad wrote Halden’s public apology for an AI mistake. Kenji Arai has been on call for an agent used by two million people. Daniel Reis runs a platform where customers increasingly let automated tools deploy code to production.

The panel takes on the practical version of the question: what an agent should be allowed to do without asking, how to design undo and audit trails, what to tell customers, and how a company divides responsibility between the people who build the model, the product and the policy.

What you will take away

• A framework for deciding which actions an agent may take on its own
• Design patterns for review, undo and audit that hold up during a real incident
• How to share accountability across engineering, product and leadership