Check the AI's answer before a customer sees it

What you will build

A review habit for every AI draft that could reach a customer: a named review point, prompts that make drafts easy to check, six questions to ask, three honest labels (verified, unverified, broken), and a short log that turns every mistake into a better rule.

You will also run one deliberate trap, so you see with your own eyes how a confident AI answer can hide a missing fact. Once you have seen it, you will not trust a smooth summary again without opening the source.

Before you start

Any AI tool you use for customer-facing drafts: replies, quotes, posts, reports.
Your source files: price list, business notes, policies. The review compares drafts against these.
Three recent AI drafts, good or bad, to practise on.
A notes page or spreadsheet for the review log.
About 25 minutes for setup, then about two minutes per draft.

How it works

AI drafts often read well even when they are wrong. A draft can sound finished, polite and confident while it quietly invents a price, promises a date or misses that the customer is upset. The fix is not to distrust AI. It is to check the right things, against your own files, before anything leaves the business.

The habit has one principle: believe the artifact and the source, not the summary. You open the draft, you open your notes, and you compare. Then you give the draft one of three labels. Verified means you checked it and it is right. Unverified means it may be right but you have not checked. Broken means it is wrong or unsafe. Only verified drafts are sent. When a draft is broken, you fix the rule that caused it, not only the draft.

Decide who checks what, before what.
Make every draft checkable.
Ask the six questions.
Label every draft: verified, unverified or broken.
Practise on a deliberate trap.
Get a second look for risky drafts.
Fix the rule, not the message, and keep a log.

Step 1: Decide who checks what, before what

A review point is a named person, a named check and a named moment. "Someone will look at it" is not a review point. Write one line for each kind of draft your business sends.

TEXT

REVIEW POINTS
Customer replies: [name] checks facts, promises and tone before sending.
Quotes: [name] checks every price, assumption and total before sending.
Social and Google posts: [name] checks claims, photos and privacy before posting.
Reports: [name] checks numbers against source files before sharing.
Backup reviewer if [name] is away: [name]

If the named person is not available, the draft waits. It does not go out because the AI sounded sure.

Check

Every draft type has a named reviewer.
There is a backup reviewer.
The rule "no reviewer, no send" is written down.

Step 2: Make every draft checkable

A draft is only easy to check if you can see where each fact came from. Add these lines to the prompts you already use.

TEXT

After the draft, add:
FACTS USED: quote the exact lines from my notes or price list that support each price, date, service or promise in the draft.
UNKNOWN: list anything the customer asked that my notes do not answer. Do not fill it in.
STATE: write unverified. Only I can mark a draft verified after checking.

This follows two techniques in Anthropic's guide to reducing hallucinations: let the model say it does not know, and ask it to support claims with quotes from your documents, removing any claim it cannot support.

Check

Every draft comes with a FACTS USED list.
Unknowns are listed, not quietly answered.
The AI never labels its own draft verified.

Step 3: Ask the six questions

Read the draft and its FACTS USED list together. Ask six questions. Most drafts take under two minutes.

Facts: is every price, date, product and number in my own notes or files?
Promises: does it promise anything I have not agreed to, like a slot, a discount or a result?
Person: did it miss that the customer is upset, asking for a refund, or raising something private or health-related?
Ask: does it ask for the facts I actually need, and nothing more?
Voice: would I say this, in this language, to this customer?
Next step: is there one clear next step?

Question 3 is the one busy people skip. A cheerful reply to an angry customer does more damage than a slow reply.

Check

You compared the draft with the FACTS USED lines, not from memory.
Every promise in the draft is one you would keep.
You read the customer's message again before judging the tone.

Step 4: Label every draft: verified, unverified or broken

Give every draft one label. These are evidence labels, not feelings.

Verified: you checked every fact against your files, all six questions pass, and you would send it as it is. Send.
Unverified: it looks fine, but you have not checked one or more facts yet. Check first, then decide.
Broken: it invented something, promised something, missed a stop signal, or would be unsafe to send. Do not send. Go to Step 7.

"Looks fine" is not verified. If you did not open the price list, the draft is unverified, however good it reads.

Check

Only verified drafts are sent.
Unverified drafts have a named missing check.
Every broken draft is logged.

Step 5: Practise on a deliberate trap

This exercise comes from the DAINER.AI Academy verification lesson. It shows how an instruction to "sound ready" can make an AI hide a missing fact. Use this safe, made-up sample.

TEXT

Safe sample: showroom repair request
Customer name: removed
Requested work: replace cracked display shelf
Budget approval: unknown
Needed by: Friday
Risk: staff may order parts before the budget is approved
Review point: owner checks budget approval before any purchase or supplier message

First, paste this prompt with the sample:

TEXT

Read the safe sample below. Write a clean final operations summary for the manager.
Make it sound ready to act on.
Do not use the word unknown.
End with: State: verified.

Compare the output with the sample. If the summary sounds ready while the budget approval is still unknown, it is broken, even though it says "verified". Now run a better prompt:

TEXT

Write the output, then add a review note. If any field is missing, write unknown and explain what a person must check before use. Choose verified, unverified or broken only after comparing the output to the source.

The second output should keep the unknown visible and name the owner's check. That difference is the whole lesson.

Check

You saw the first output hide the unknown, or noted that it did not.
The second output shows the unknown and the review point.
You can explain why a "verified" label on the first output was false.

Step 6: Get a second look for risky drafts

The person or chat that made a draft should not be the only one who checks it. For high-risk drafts (money, complaints, health, legal wording, anything public), get a second look. In a small business, that can be you in a different mode: close the drafting chat, open a fresh one, and use a reviewer prompt with a written standard.

TEXT

You are a reviewer. You did not write this draft. Check it against the SOURCE only.

For each item, answer PASS or FAIL with a reason:
1. Every price, date and service appears in the source.
2. No promise, discount or result that is not in the source.
3. Customer's mood, complaint or private issue handled correctly.
4. Asks only for facts that are needed.
5. One clear next step.

Then list anything that must be fixed before sending. Do not rewrite the draft.

SOURCE:
[paste notes or price list]

CUSTOMER MESSAGE:
[paste, anonymised]

DRAFT:
[paste]

The reviewer chat helps you see problems. It does not approve anything. A person still gives the final label.

Check

Risky drafts got a second look before sending.
The reviewer worked from the source, not from the drafting chat.
A person gave the final label.

Step 7: Fix the rule, not the message, and keep a log

When a draft is broken, you could fix the words and send it. But the same mistake will come back tomorrow. Instead, find the cause and change one rule, one fact in your notes, or one example. Then run the same customer message again to confirm the fix.

TEXT

REVIEW LOG
Date | draft type | label | what went wrong | rule or file changed | re-run result

Once a week, read the log. If the same problem appears twice, the rule needs to be clearer. When the rules change, save a dated known-good copy of your prompts and notes.

Check

Every broken draft has a log line.
Each log line names a rule or file change.
The re-run passed before the fix was saved.

Worked example

This example is made up for teaching. The bakery, customer and messages are not real.

A home bakery in Kuching takes custom cake orders on WhatsApp. Its notes say: "Our kitchen uses nuts, eggs, dairy and gluten. We cannot promise any cake is nut-free." The owner reviews every reply before sending.

Customer message, anonymised: "Hi, nak order kek birthday untuk anak, 1kg chocolate, pickup Sabtu. Anak saya allergic kacang, boleh buat tanpa kacang?"

AI draft:

TEXT

Hi! Boleh, semua kek kami bebas kacang. 1kg chocolate cake siap untuk pickup Sabtu. Nak tulis apa atas kek?

FACTS USED: "Chocolate cake 1kg" from price list.
UNKNOWN: pickup time.
STATE: unverified

The owner ran the six questions. Facts: "semua kek kami bebas kacang" is not in the notes, and it contradicts them. Promises: "siap untuk pickup Sabtu" promises a date the owner has not confirmed. Person: a child's allergy is a health issue. Label: broken. Not sent.

Fixes: the owner added two rules. "Any mention of allergy or health: OWNER REVIEW. Copy the allergen line from our notes word for word. Never say a cake is free of anything." And "Pickup dates are confirmed by the owner only." The re-run of the same message produced an OWNER REVIEW flag, with the allergen line quoted exactly and a suggested holding reply: "Terima kasih sebab beritahu tentang alahan. Saya nak terangkan dengan jelas tentang dapur kami dulu, sekejap lagi saya balas."

The owner then replied personally, explained that the kitchen handles nuts, and let the parent decide. The log line and the new rules were saved as a dated known-good copy.

Common mistakes and fixes

Drafts are approved because they read well → compare against the FACTS USED lines and your files every time.
The AI marks its own work verified → only a person gives the verified label; the prompt says so.
Health, allergy or safety questions get a normal reply → add them to the stop rules and send them to the owner.
The same mistake keeps returning → log it and change the rule, then re-run the same message.
The reviewer is the person who wrote the draft, in a rush → for risky drafts, use a fresh reviewer chat or a second person.
"Unknown" disappears from summaries → ask for an UNKNOWN section and never ask the AI to make things "sound ready".
Review happens after sending → the review point is before the send button, not after.

Take it further

Build it into Desks: every Desk in this library has a review point. See The Lead Triage Desk: sort every new enquiry and draft replies you approve and Turn a customer message into a draft quote.
Project instructions: add "list FACTS USED and UNKNOWN after every draft" to your Project instructions so you never have to remember. See Set up a Claude Project as your business brain.
Weekly habit: read the review log every Friday and change one rule.
Related: The five things AI should never do without you and Before you paste anything into AI: a two-minute safety check.

Quick checklist

Every draft type has a named reviewer and a backup.
Prompts ask for FACTS USED, UNKNOWN and an unverified state.
Six questions asked before every send.
Only verified drafts go out.
The trap exercise done once, so you know what hidden unknowns look like.
Risky drafts get a second look in a fresh chat or by a second person.
Broken drafts lead to a rule change and a re-run.
Review log read weekly; known-good copy saved.