What AI must not decide, and how to double-check an answer

What you need

• Any AI chat assistant and a shared document
• A search assistant that shows its sources
• The official websites of the bodies that set the rules in your sector
• Gamma for the privacy sheet
• A booking page or the contact details of your professional advisers

1. Write the red-line list

   List the decisions in your business where a wrong answer costs money, health or legal trouble: tax, contracts, employment, safety, medical questions, large payments. AI can help you prepare, but these decisions stay with you or a professional.

   Example

   Red line: AI may explain terms and list questions, but does not decide on tax filings, contract terms, dismissals, safety rules, medical questions or payments above an amount I set.

2. Double-check in three steps

   One: ask the assistant for its sources. Two: open each source and find the exact sentence. Three: confirm it on the official website of the body that sets the rule. If the three do not match, the answer is not ready. For checking figures and statistics, see /guides/check-a-number-before-you-publish/.

   Weak

   Is this right.

   Better

   Give me the source for each claim in your answer, with the exact sentence you relied on. Mark any claim you cannot source with [NO SOURCE?].

3. Ask five viewpoints to attack the answer

   Joshua shared a council idea: instead of one reply, five viewpoints look at the problem. One looks for what will fail, one rebuilds the idea from scratch, one looks for the upside you are missing, one ignores all context and one cares only about the next action. A chair then sums up. You can run the same thing as one prompt in any AI chat assistant.

   Example

   Act as a council of five advisers and check the answer below before I act on it: the contrarian (what will fail), the first-principles thinker (rebuild it from scratch), the expansionist (the upside I am missing), the outsider (ignore all context) and the executor (one action for today). Then, as the chair, list every claim the answer depends on with fact or guess and its original source, and give one next step.

   • Source: Joshua (@nocode.joshua) on Instagram

4. Split facts from guesses and trace each claim

   Mariah Brunner lists rules for better AI work, among them: make it challenge your assumptions, separate facts from assumptions and trace important claims back to the original source. Ask for a table of every claim with fact or guess and where it comes from, then open the sources yourself.

   Example

   List every claim in your answer in a table: the claim, fact or guess, and the original source. If there is no source, write NO SOURCE.

   • Source: Mariah Brunner (@itsmariahbrunner) on Instagram

5. Make a privacy do and do-not sheet

   One page that says what may go into an AI chat and what may not, with two or three examples from your own work. Keep it next to the red-line list.

6. Prepare questions for a professional

   When something is on the red-line list, use AI to sort your notes and write clear questions, then take them to an accountant, lawyer or doctor. This guide gives no medical, legal or financial advice.

7. Decide, and keep a one-page check note

   Before you ask, write one line about what you will decide or publish because of the answer; it tells you how carefully to check. Take the one next step only after the claims you rely on are checked. Keep a short note: the question, what you checked, what was wrong and what you decided. Next time a similar answer comes up, you start from that note.

Copy it into any AI chat assistant and change only the first line.

Change only this line: BUSINESS = [a home cleaning service with five staff]

You help me write a red-line list of decisions that AI must not make for my BUSINESS, and a routine to double-check AI answers.

First ask me, one question at a time: which decisions I make often (for example pricing, refunds, hiring, contracts, tax) and which people I can ask (for example an accountant, a lawyer, a safety adviser).

Do three things:
1. Sort my decisions into two lists: AI may help prepare, and a person or professional decides.
2. Write a three-step double-check routine in plain words: ask for sources, open them, confirm on the official website.
3. For one red-line decision I name, write five clear questions to take to the right professional.

Give no medical, legal or financial advice yourself. Where something depends on law or health, write [CHECK WITH A PROFESSIONAL?].

After my answers, start with the two lists.

Copy the instruction

A clear red-line list, a three-step check you run before acting on any AI answer, a privacy sheet and better questions for the professionals who decide.

• Acting on an AI answer about tax, law or health without asking a professional.
• Accepting a source without opening it.
• Treating a confident tone as proof.
• Running the check after you have already acted on the answer.
• Treating the council's view as proof: it is still AI, the sources are the proof.