A runbook explains how one AI-assisted task is completed and what happens when it fails. Choose the task using the AI assistance and accountability guide. This example uses a draft customer response; adapt the checks to the actual workflow.
Define the approved input
Record the sources that may be used: for example, an authorised order-status record and the current delivery policy. Specify which details are necessary and which must be excluded. Identify the owner who updates the policy and how the workflow recognises an outdated version.
Do not copy an entire customer history into the tool when the task only needs an order reference and current status. Confirm the tool’s approved handling of information before processing confidential data.
Make the output contract explicit
The draft should answer the customer’s question, use only supported facts and identify anything that requires a human decision. Require an escalation when the order record is missing or contradicts the policy. An instruction to guess a plausible answer defeats the purpose of the workflow.
For this illustrative task, the tool prepares a draft; an authorised person decides whether it can be sent. The business’s existing customer commitment process remains part of the runbook.
Check the result against the source
- Does the response concern the correct order?
- Are dates, amounts and policy statements supported?
- Does it make an unauthorised promise or conceal uncertainty?
- Is the requested escalation visible?
- Has the reviewer recorded the outcome?
The NIST AI framework provides wider voluntary risk-management guidance. These checks are an operational example, not proof of conformance to a standard.
Maintain an exception queue
Record the task reference, failure category, responsible person and next action. Use protected references rather than copying sensitive content into a broad error log. Keep unresolved work pending; a generated draft does not mean that the customer received a verified answer.
Test the manual fallback
Have another authorised person complete the task without the AI tool, using the documented sources. Verify that the task can continue during an outage or after an unacceptable output. Record who may pause the AI step and how pending items are recovered.
Review changes and performance
Compare completion, corrections, escalations and total handling time with the previous process. Retest representative failures after significant changes to the tool, inputs or instructions. Expand the workflow only when the evidence supports that next scope and the required reviewers can handle it.
