Imagine asking AI to write a reply about your service. It produces a clear, polite email, then adds: “We can deliver tomorrow.”
You never supplied that promise.
This is an illustrative example, but the underlying problem is real: generative AI can produce incorrect information confidently.
By the end of this issue, you will have three checks to use before an AI-assisted message reaches a customer: verify the facts, spot added promises and identify the person who can approve it.
In issue one, we looked at what information staff can put into AI. This time, we are checking what comes back.
1. CHECK THE FACTS AGAINST THE SOURCE
Look for dates, figures, names, links and claims about your service. Check them against your approved service information or the original document. If AI supplies a reference, open it and check that it supports the claim.
ILLUSTRATIVE DEMONSTRATION
Approved website text: “We provide weekday deliveries. Delivery dates are confirmed when you book.”
AI draft: “We deliver seven days a week.”
The draft has changed the service. A corrected version would be: “We provide weekday deliveries. We will confirm your delivery date when you book.”
The source gives you something concrete to check. A polished sentence is not evidence that its facts are right.
2. SPOT WHAT IT HAS ADDED OR LEFT OUT
Compare the answer with the task you gave it. Has it invented a delivery time, removed a condition or turned a suggestion into a promise?
ILLUSTRATIVE DEMONSTRATION
Invented approved service text: “We offer a free initial consultation of up to 20 minutes. Further work is quoted separately.”
Illustrative AI draft: “Your consultation and follow-up work are free.”
A corrected version would be: “We offer a free initial consultation of up to 20 minutes. Further work is quoted separately.”
The draft removed the time limit and extended the offer to paid work. Restore both conditions before using it. An accurate summary must preserve the limits of the original offer.
A fluent answer may need substantial correction. If you cannot verify an important claim, leave it out or hold the draft until you can.
3. CONFIRM WHO CAN APPROVE THE MESSAGE
The reviewer needs to understand the task and have access to the information needed to check it. For a routine message, that may be the person sending it. For a price, deadline or service commitment, involve whoever has authority to agree it.
ILLUSTRATIVE DEMONSTRATION
AI draft: “We can offer you a 10% discount.”
If no approved offer supports that sentence, remove it. If you want to make an offer, ask the person authorised to agree prices before including it.
These checks help you review a draft. They do not guarantee that every error will be caught.
Evidence: NIST’s Generative AI Profile (2024), section 2.2, describes confidently incorrect outputs. Action MS-2.5-003 recommends reviewing generated sources and citations. The three checks here are our practical application of that evidence; the demonstrations are invented examples.
THE TAKEAWAY
Before sending an AI-assisted message, check its facts, check for added promises and confirm who can approve it.
ONE ACTION
Review one AI-assisted draft using the three checks above before sending it. Use an approved tool and information already published on your website; keep personal and confidential information out.
Record unsupported claims, missing conditions and corrections. Add the minutes spent drafting, checking and correcting, then compare the total with doing the task normally.
KEEP / ADAPT / REJECT
Keep the task if its output can be checked reliably and it saves time after review. Adapt the instructions or reduce the task’s scope if corrections remove the benefit. Reject this use of AI if important claims cannot be verified or checking and rework outweigh the saving. One successful draft is a trial result, not proof of reliability.
Barbara
AI Made Human
P.S. Keep the three checks beside you when you review your next draft: facts, promises, approval.