Incident Response for AI: What to Do When a Model Fails

This article outlines what an AI incident response plan should contain: detection and triage, the authority to pause or roll back, communication, and a blameless review that turns each failure into a stronger system.

The ProblemFailure without a plan

Every deployed system fails eventually, and AI systems are no exception; they produce a harmful output, drift out of spec, or behave unexpectedly under conditions no one anticipated. The damage from such an event is shaped less by the failure itself than by how the organization responds. Too often there is no plan: the failure is met with confusion, ad hoc decisions, and delay, while the harm spreads and the explanation arrives late and incomplete. Improvising the response in the middle of an incident is the worst time to design it.

Why It MattersThe response is what protects people

When a model fails in a high-stakes setting, the people affected depend on how quickly the problem is caught, contained, and communicated. A prepared organization can detect the issue early, limit its reach, tell the affected parties honestly, and correct course. An unprepared one lets the harm compound while it works out who is responsible and what to do. The difference is not luck; it is whether the response was planned before it was needed. Incident response is, in the end, a form of care for the people on the receiving end of a system's mistakes.

The TeraSystemsAI PerspectivePlan the response before the failure

We hold that any serious deployment must come with an incident plan written while no one is panicking. That plan names who is responsible, defines how a failure is detected and escalated, and sets out the steps to contain it and communicate about it. It treats failures as expected events to be managed, not as surprises to be denied. And it includes the discipline of learning afterward, so that each incident strengthens the system rather than merely embarrassing it. Preparing for failure is not pessimism; it is the basic responsibility of deploying something that can affect people.

Practical ImplicationsA response that works under pressure

In practice, incident response means defining detection and escalation in advance, so a problem reaches the right people fast, with the authority to act. It means having containment options ready, the ability to pause, roll back, or fall back to a safe mode, rather than inventing them mid-crisis. It means communicating honestly and promptly with those affected, since silence erodes trust faster than the failure itself. And it means running blameless postmortems that focus on what in the system allowed the failure, not on whom to blame, so the lessons actually take hold. A system is only as trustworthy as the plan for what happens when it goes wrong.