Your AI document tool stops responding at 9 a.m. Staff can still open customer records, but they cannot prepare the summaries used for the afternoon service calls. A retry loop will not tell the manager which commitments to protect.
Plan the business fallback before an outage. Decide which work can continue, which work should wait, and who can change the operating mode. The fallback must fit the people available that day.
NIST's Generative AI Risk Management Profile provides voluntary guidance for managing AI risks. It does not prescribe a universal outage duration or recovery guarantee. Your continuity plan needs to reflect the actual work and dependencies.
Run a thirty-minute tabletop
Use this fictional exercise with the operating manager, a staff member who does the work, and the person responsible for the tool. Tell them the document tool is unavailable for the rest of the morning, while the records system remains available.
| Exercise minute | Question | Decision to record |
|---|---|---|
| 0–5 | Which work is already waiting or partly complete? | Identify afternoon commitments and unfinished summaries |
| 5–12 | What can staff complete manually with available access? | Assign priority customers to a manual preparation queue |
| 12–20 | What will not fit before the customer deadline? | Choose postponements and the person who communicates them |
| 20–25 | How will staff avoid duplicate work when service returns? | Track which records they completed during the outage |
| 25–30 | What conditions justify restart? | Check one pending item and resume in a controlled order |
The exercise lasts thirty minutes; it does not promise a thirty-minute repair. Use the time to expose missing access, unclear priorities, and a fallback that exists only on paper.
Make the manual queue small enough to work
In the example, twelve summaries are due before calls. Two staff members can prepare six in the available time without abandoning other committed work. The manager chooses the six customers whose calls require the summary and reschedules the remaining calls with a truthful explanation.
Those numbers are hypothetical capacity assumptions. The important step is choosing the customer commitment you can honor. Telling staff to “do everything manually” leaves them to make priorities under pressure.
Have one person demonstrate the manual preparation using a permitted record. They need access to the underlying facts, a usable output format, and a place to record completion. If the only copy of the facts sits inside the unavailable vendor, the fallback requires a different design.
Keep outage work distinct from pending automated work
Mark records completed by staff and record items they postponed. When the vendor returns, the operating team should know which jobs still need processing. Review queued submissions before replaying them so a restored tool does not repeat customer messages or overwrite staff work.
The implementation team can design the technical controls. The business owner must define what counts as completed and which customer-facing action requires another check. A status page turning green does not answer those operational questions.
Rehearse a restart choice
Resume one representative pending item, inspect its result, and confirm the staff-completed records remain intact. Then release the rest in an order the review team can handle. If the source records changed during the outage, account for those changes before processing the old queue.
Keep incorrect-answer containment and routine maintenance in their own plans. Our maintenance guide explains support scope; The Ops Guide's fractional CAIO services can help leaders agree on a workable continuity decision across teams.

