Rehearse manual recovery before AI routes requests

Suppose an AI tool routes maintenance requests and starts sending urgent cases to the routine queue. Turning it off may stop the next assignment. The previous requests can still be sitting with the wrong team, and a tenant who already reported a fault may have no reason to call again. A recovery drill should show that staff can stop new errors and find the people affected by the earlier ones.

The US National Institute of Standards and Technology’s AI risk management playbook recommends planning how to bypass or deactivate an AI system while limiting disruption to service, including backup procedures and criteria for bringing the system back. Its generative AI profile also advises setting criteria for deactivation and planning how to tell affected users about workarounds.

Check what actually stops

Use a test environment with dummy records and notifications routed to a test inbox. Put one request in each state the team expects to see: waiting, in progress and already assigned. Mark one as deliberately misrouted. Give a shift lead a clear stop trigger, such as a confirmed urgent request appearing in the routine queue, and a way to pause routing without asking the model.

After the lead pauses routing, enter another dummy request. Check the queue and the receiving system for fresh automated assignments. A paused screen does not prove that background jobs, retries or messages already sent to another system have stopped. For each action still in flight, identify who can cancel it and how the team will confirm what happened.

Account for every request before restarting

Have a named colleague receive new requests through the manual route during the drill. Compare the original intake records with the resulting assignments. Account for each request as completed, pending with a named owner or corrected. If someone has already received a wrong assignment notice, send a correction through the test channel; restoring an earlier system version will not recall a message.

Set a maximum time to resume manual intake based on how long people can reasonably wait for the service. Measure from the stop trigger to the first correctly logged manual request. Record the elapsed time, any unaccounted records and any duplicate assignments. If requests go missing or get assigned twice, fix the cause and rerun the case before relying on the handover. Require the service owner to check corrected routing on dummy cases and approve restart once the pending work is accounted for.

Keep the handover instructions somewhere the next shift can reach even if the AI dashboard is down. Pair them with the limits you set on what the tool can do.

Sources and further reading

• NIST AI Risk Management Framework Playbook: Manage
• NIST AI 600-1: Generative Artificial Intelligence Profile, July 2024

Sources checked 4 November 2026. Prepared with AI assistance and editorial review.