A One-Page Incident Plan for AI Gone Wrong

You don't need a fifty-page incident response manual. You need one page with five boxes, and a human name in every box.

Box one: Detect. How will we know something went wrong? A customer complaint, a staff member spotting a bad output, an automated alert. Write down the actual signals, not "we'll notice."

Box two: Contain. Who can pause or disconnect the tool, and how? If the answer requires three approvals, your incident will spread while the paperwork moves.

Box three: Assess. Who figures out what happened, who was affected, and what data was involved? Name the person who leads this, not "the team."

Box four: Notify. Who tells leadership, who tells affected customers, who talks to the vendor? Decide the order now, when nobody's panicking.

Box five: Learn. Who writes down what happened within a week, and who makes sure the fix actually gets implemented?

Don't automate chaos. A fast AI tool with no incident plan just fails faster. One page, five boxes, reviewed twice a year. That's the whole thing.

I explain frameworks; I don't give legal counsel.