What should your AI workflow do when a source or system is unavailable?

Distinguish missing evidence from an ordinary blank

A required field can be missing because the requester never supplied it or because the source cannot be reached. Those need different next actions. The intake owner asks for an absent contact; the system owner investigates an unavailable source. Record which condition applies so the work does not cycle through the wrong queue.

Do not serve an old answer as confirmed current guidance simply because it is available. Your procedure owner defines when a previous approved version may be used and for which task. If that permission is not established, hold the request and show the unavailable source. A confident guess is not continuity planning.

Make status tell the truth

Prepared means the draft exists. Attempted means an action was requested. Confirmed means the destination supplied the agreed evidence. Keep those states separate. A network timeout can leave the action outcome uncertain, so neither success nor failure should be invented.

Give the responsible person the request reference, attempted action, last confirmed event and next check. Avoid unnecessary private payloads in the incident record. The person needs enough evidence to recover the task without turning a diagnostic log into another unrestricted data store.

Prevent a retry from becoming duplicate work

Agree a stable reference and a way to check the destination for an existing result. If an action may already have succeeded, do not blindly repeat it. The system owner defines the recovery behavior for the exact application and action. A generic retry button is not enough when creating another record can create another job.

Show the operating team what it can do manually without creating a second request. Preserve the original reference and acceptance history. If the manual action is taken, record it so the restored workflow does not later repeat it as though nothing happened.

Illustrative editorial photograph, not a customer result.

Exercise the fallback in a controlled test

Use a permitted test environment or separately approved outage simulation. Define the expected hold, notification and manual step before the exercise. Do not disrupt a live source merely to test a marketing demonstration. The appropriate system and process owners authorize the method and review its result.

Ask the receiving employee to continue using the fallback. Record whether they can find the current procedure, identify the correct owner and avoid an unauthorized action. A document describing recovery is not evidence that the team can actually use it.

Measure the broken day in the pilot

Include time spent detecting, reviewing and recovering failed requests. Keep unconfirmed and held work in the sample. A trial that counts only successful drafts will understate the operating burden. Record how the failure affected waiting separately from staff labor time.

Use the same acceptance rule after recovery. A task should not be marked complete because support says the connection is restored. Inspect whether the destination accepted the original request and whether any manual duplicate remains. Close the incident with evidence the process owner can understand.

Add a recovery record before rollout

Put source failure, destination failure, uncertain outcome, manual fallback, owner and next check into the Operations AI Pilot Charter. Pair it with the connection guide to define what confirmed means for your actual tools.

NIST's AI Risk Management Framework offers broader voluntary risk-management context; it does not certify this worksheet or a product. Request a relevant demo and ask our team to show a failed-source case alongside the normal workflow. SimplSolutions confirms the manual path, exact actions and ongoing owners in the proposed scope.