LESSON 5 OF 6
Keep it accurate and on-brand
BY THE END OF THIS LESSON
Put a review process around AI output before it reaches customers.
The risk is quiet, not loud
The dangerous failure is not obviously broken output — someone catches that. It is a plausible, well-written message containing a policy you do not have, a figure that is slightly wrong, or a commitment you cannot keep. It reads fine. It ships.
Three guardrails worth the effort
A claims allowlist. The specific facts and figures anything customer-facing may state. Anything outside it needs a human.
A named reviewer. One person, by name, for each output type. "The team reviews it" means nobody does.
A visible confidence flag. Instruct it to mark anything it is unsure about in brackets, so the reviewer knows where to look instead of re-reading everything.
Review the process, not just the output
Once a month, look at what got through. You are not looking for typos — you are looking for the category of mistake that keeps appearing. That tells you which instruction is missing, and fixing it removes a whole class of errors rather than one instance.
WORKED EXAMPLE
UNBOUNDED
Write a reply explaining our refund policy to this customer.
BOUNDED
Write a reply using only these approved facts: - 30-day refund window from purchase date - Refunds go to the original payment method, 5–10 working days - Annual plans are pro-rated; monthly plans are not If the customer's situation is not covered by exactly one of these, do not answer — write "[NEEDS REVIEW: not covered by policy]" and stop. Mark anything you are less than certain about in square brackets.
The first invents a plausible policy. The second physically cannot, and it tells the reviewer precisely where to look rather than making them re-read the whole message.
YOUR TURN
Build your claims allowlist.
RUN THIS
Help me build an approved-claims list for customer-facing AI output. Ask me for the facts, figures, policies and guarantees we are willing to state publicly. Push me for exact numbers rather than approximations. Then produce: the allowlist as a block I can paste into prompts, plus a list of the questions most likely to fall outside it — the ones I need a human answer ready for.
What a good result looks like
The second list is the valuable one. Those are the questions that will otherwise get an invented answer.
KNOWLEDGE CHECK
Answer all 3 correctly to complete this lesson.
1. What is the most dangerous kind of AI output error in a business?
2. Why name a specific reviewer rather than "the team"?
3. What should a monthly review of AI output look for?
REMEMBER THIS
Allowlist the claims, name one reviewer, and make uncertainty visible.