AI adoption
Before you send the AI draft, make the check specific.
Give reviewers a source, clear checks and permission to stop—not just an instruction to look over the draft.
An AI-drafted customer reply is ready to send. It reads well, sounds helpful and includes a delivery date. The person reviewing it now needs to establish whether that date was confirmed or simply appeared in the draft. “Someone checks it” is not a complete working instruction.
Define what the reviewer must establish
Start with the mistakes that would matter in this particular task. For a customer reply, that might mean the wrong recipient, an unsupported commitment, an incorrect amount or a missing qualification. Checking tone and spelling does not settle those questions.
Write a short checklist next to the task, rather than burying it in a general AI policy. Make it specific enough that two colleagues could explain what they are checking.
- Match names, dates and amounts to the relevant source.
- Check that any commitment has been authorised.
- Look for important conditions or omissions, not just incorrect statements.
- Confirm that the content is appropriate for the intended recipient.
Put the evidence beside the draft
The reviewer needs access to the current information that can support or contradict the answer. A plausible explanation from the same AI tool is not a substitute for that source. Where the draft includes links or quotations, open the original material and check what it actually supports.
Consider a hypothetical service business drafting an update about an installation. The job record says a part is awaited; the draft says the installation will happen on Thursday. The check is not whether Thursday sounds reasonable. It is whether someone authorised to make that commitment has confirmed it.
Use only information approved for the chosen tool. If the necessary evidence cannot be used there, keep the comparison in the existing authorised system or handle the task manually.
Give the reviewer a real stop option
Name a reviewer who understands the task and has enough time and authority to reject the output. Agree what happens when evidence is missing or contradictory: hold the message, ask the responsible colleague, or return to the normal manual process.
For this kind of customer-facing workflow, keep sending separate from drafting until your checks and responsibilities are established. An approval button should represent an actual decision, not a required click on the way to finishing the queue.
Learn from corrections, not just approvals
Keep a small record of the types of correction made, without unnecessarily copying customer information into another log. Repeated invented commitments or missed conditions are reasons to revisit the workflow, instructions or suitability of the task.
Walk through a few synthetic examples together, including one that must not be sent. Review the checklist when the process or source information changes. This makes checking part of ongoing adoption, rather than a one-off training slide.
Rethink’s AI working sessions help teams turn these decisions into practical ways of working, with further adoption support scoped by agreement. Start by making one existing check precise enough for a colleague to use.