AI can turn a rough customer email into something calm, clear and professional. It can also make an unsupported promise sound like it came straight from the managing director. Lovely sentence. Potentially expensive Tuesday.
The problem is easy to miss: we read for tone and accidentally give the facts a free pass. A sentence can be well written and still be wrong. NIST identifies confidently presented false information as a generative AI risk. That is why the final check needs more than “Does this sound good?”
Try this on a fictional example
Imagine a sign shop preparing a reply about a customer order. The approved notes say: “We are checking stock. Installation date is not confirmed.” A polished draft instead says: “Your installation is scheduled for Friday.”
That is a made-up example, not a client case study or a recorded model test. The rewrite added a commitment that the source never made. The customer will remember Friday. They will not care which software supplied it.
Separate the writing from the verification
First, give the tool the facts it may use. Start with fictional information or material permitted in your organization’s approved system. Ask for the tone and structure you need, but tell it to flag gaps rather than fill them.
Then check three things yourself:
Facts: names, figures, dates, product details and anything presented as certain.
Promises: delivery times, discounts, guarantees, refunds and next steps someone now has to honor.
Permission: whether the tool was allowed to receive the information and whether the recipient should receive the finished message.
A prompt worth keeping
Rewrite the notes below as a clear, friendly customer email. Use only the facts supplied. Preserve uncertainty. Do not invent dates, prices, guarantees, commitments or policy details. If a fact is missing, place [CONFIRM] in the draft. After the email, list every claim or commitment that needs human verification. Notes: [insert approved or fictional notes].
Those instructions are helpful constraints, not a guarantee. Compare the result with the actual source. Asking the same AI whether it is correct is not an independent fact check. A named person still owns the decision to send.
The small habit that changes the workflow
On the next low-risk draft, run two separate passes: one for clarity, one for accuracy. Keep the original notes beside the final message. If the draft changes what the organization is promising, stop and resolve it before sending.
AI can help you write the email. You still own what it promises.
Your turn
What would be the most costly detail for an email to get wrong in your workplace? Keep examples general and leave out customer or employee details.
Join the conversation on Bluesky · ← Back to Reasonably Intelligent
