Operations
Respond to AI failures as product incidents
When an AI feature fails, contain the impact, preserve useful evidence, and separate incorrect output from application errors before deciding how to recover.
An AI incident may look different from an ordinary outage. The application can remain available while producing misleading summaries, retrieving the wrong records, or repeating an action. Imagine a hypothetical account assistant that begins attaching unrelated source references to answers. The first objective is to contain the impact and understand affected work. Debating whether the model or the code deserves blame does not help users recover.
Classify the visible impact
Describe what users are experiencing in concrete terms. Is the feature unavailable, slow, incorrect, exposing unauthorized information, or executing an unintended operation? Determine which workspaces, task types, and release revisions may be affected. Distinguish confirmed evidence from hypotheses. Assign an incident owner and a place for the working timeline. That shared record reduces duplicate investigation and keeps configuration changes from becoming undocumented experiments during an already confusing situation.
Contain the problem proportionately
Choose the narrowest control that reliably stops the harmful behavior. It may be disabling one tool, pausing a queue, reverting a prompt bundle, or taking the feature out of service. Preserve a usable fallback if one exists. Consider tasks already in progress and outputs already saved. In the source-reference example, holding new answers for review may be appropriate while the team identifies whether previously published answers also need correction.
Collect evidence without expanding exposure
Capture task identifiers, relevant revisions, timing, error categories, and a small set of authorized examples. Avoid copying large amounts of private content into an open incident channel. Trace the path through retrieval, model response, validation, and persistence. Reproduce the issue with synthetic material when possible. An apparently incorrect model answer may originate from stale source selection, an adapter bug, or a user interface that displays the wrong saved result.
Recover and prevent recurrence
Verify the correction against the incident examples and adjacent cases before restoring normal operation. Check that affected records or outputs receive the appropriate review. Communicate known impact and recovery status through the established product process. Afterward, write a concise account of the trigger, contributing conditions, detection gap, and follow-up owners. Add a meaningful regression test or operational check, and remove temporary diagnostics or permissions introduced during the investigation.
