KINSMC
A community workshop uses an AI tool to turn registration requests into draft schedule confirmations. On Monday, a coordinator notices that a draft promises a Wednesday slot that the booking calendar shows as full. There are eleven other drafts in the queue and three confirmations already sent. Closing the AI tool would stop new drafts, but the team would still need to decide what to trust, what to check and who should contact the people already booked.
A manual fallback is the short procedure that answers those questions. It identifies when to stop using AI, where the authoritative records live and how to finish the task by hand. It also gives the person who finds the problem enough authority to act while someone else investigates. This guide uses a fictional workshop so you can adapt the procedure to your own routine.
Write a stop rule that a colleague can apply
Start with the worst plausible mistake in the task. For workshop scheduling, sending the wrong day or promising a full session matters more than a stiff sentence. A rule for the team could say that one confirmed wrong booking or one inaccessible booking calendar pauses all AI-generated confirmations. Spelling and tone fixes can go through ordinary review. If several seemingly unrelated drafts repeat the same error, treat the batch as suspect while you work out whether they share an input or template.
Name the person on duty who can pause the workflow and a backup who can do it in their absence. Tell them which action actually prevents new work from reaching recipients. In a manual drafting setup, that may mean telling staff to stop using the draft template and holding unsent messages in a shared queue. In an automated setup, it may mean using a documented pause control and confirming whether messages already queued can still send. Write the scope in plain language, such as “pause workshop confirmations” rather than “turn off AI everywhere.”
Give the manual route a reliable starting point
Keep the source records available through the normal workspace. In the workshop example, the registration form supplies the request, the booking calendar supplies available places, and the sent folder confirms what people have received. The fallback should tell the coordinator which record wins if two sources disagree. If the calendar itself is unavailable, the coordinator can acknowledge receipt and explain that a slot will be confirmed later. They can safely keep the queue moving without inventing availability.
Save a short blank confirmation template alongside the instructions. In this example it needs only a booking reference, session, date, start time and contact route. Write down where to mark the completed task so a colleague can pick up the remaining work. During the pause, give each request a single owner and one status. Useful statuses are awaiting verification, confirmed and correction sent. Keep a reference to the source record so another person can resolve a discrepancy without searching the entire inbox.
Reconcile queued work before sending anything else
Find the last batch someone checked against the original records. Start with the requests handled since that point. Separate unsent drafts from sent confirmations using the actual sent record. For unsent work, read the original request and current calendar and write a fresh confirmation from the template. For already sent messages, verify the details and correct any affected recipient directly. Avoid sending a second confirmation to everyone just because their drafts appear in the queue.
Document the scope of the problem, the stop time, the current queue owner and the source used to verify each affected booking. The NIST Generative AI Profile discusses recovery, escalation and communicating alternative procedures when a generative AI system is disengaged. Our guide to keeping a human in the loop at work explains how to review routine outputs before they are used. The fallback starts where normal review finds a defect serious enough to interrupt the workflow.
Test the fallback and define the return to AI
Run a short drill using invented requests. Include one wrong booking that has already been sent, one duplicate request and one case with an unavailable calendar. Ask a colleague to take over using the written instructions and no coaching. Check whether they can tell which messages were actually sent, where to verify a date and who owns the queue. If they need to contact the person who designed the procedure to find a template, put the template where the team already works.
Before restarting, identify what changed, rerun the case that failed and try ordinary and awkward examples against the verified source. Ask the named owner to approve a small restart and check the first batch. The NIST AI Risk Management Framework Core includes post-deployment monitoring, override and recovery in its risk-management outcomes. If you need a way to judge whether the repaired process still saves enough time, use our guide to testing an AI workflow against the manual version. Keep the manual route usable even after the first successful batch.