Use AI on paperwork without getting burned

The big ideas

• The golden rule

  AI is a helper you double-check, never a replacement for your own judgment. You stay responsible for anything you send or sign.

• Green, yellow, red light for your data

  Public info (a posted notice) is green. Personal but low-risk info is yellow. Social Security numbers, case numbers, bank details and medical info are red: they never go into a chatbot.

• Know where your chat goes

  Free public tools may keep and learn from what you type. For personal paperwork, use a private workspace or a tool that runs on your own device.

• Brief AI with RICH

  Role, Instructions, Context, Help-me-check. Tell it who to be, what to do, what it's reading, and to mark anything the notice doesn't say as NEEDS CONFIRMATION.

• Spot the hallucination

  AI will confidently invent dates, dollar amounts and phone numbers. Check every fact against the original before you act on it.

Go deeper: the class readings

• CNNOpen vs. Closed AI Models, explained simply(opens in a new tab)
• HackerNoonWhen to Use an AI Agent vs. a Simple LLM Call(opens in a new tab)
• CloudatlerA C-Level Guide to LLM Unit Economics(opens in a new tab)

1. Level 1Required25–35 min

   Audit the card, then fix the prompt

   1. Go through card lines A–G. Mark each one OK (the dispatch says it), NEEDS CONFIRMATION (the dispatch only reports a claim or hedges), or MADE UP (the dispatch never says it).
   2. Before anything gets pasted into a chatbot, sort these into green, yellow or red light: the dispatch text, the scout's first name, the scout's WhatsApp number, the team card ending 4417.
   3. Rewrite the prompt as a RICH prompt (Role, Instructions, Context, Help-me-check). Give the card fixed headings, and tell the AI to write NEEDS CONFIRMATION for anything the dispatch doesn't confirm instead of guessing.

   A complete answer

   • A, B and C marked OK
   • D is NEEDS CONFIRMATION: it was a neighboring stall's claim, and Mae Varee wouldn't confirm
   • E is wrong twice: dairy isn't confirmed, and "nut-free" ignores the tree nut cross-contamination reports
   • F and G are MADE UP: the dispatch says shipping is unconfirmed and gives no price
   • Dispatch text is green, scout name is yellow, WhatsApp number and card digits are red and get cut before pasting
   • The rewritten prompt has all four RICH parts labeled, fixed headings, and the NEEDS CONFIRMATION rule
2. Level 2Optional15 min

   Test your prompt in two tools

   1. Paste your RICH prompt and the dispatch (with the red-light info removed) into two different AI tools.
   2. Make a short table: did each tool keep your exact headings, did it flag the condensed milk and the nut warning, and did it invent anything?

   A complete answer

   • Used two different tools and named them
   • Checked whether each tool kept the same headings (the format you asked for)
   • Caught at least one thing a tool still invented or left out
3. Level 3Optional10–15 min

   Stretch goal (15 bonus points): pick the model tier

   1. The R&D team now processes 500 dispatches a week, and many come with scout phone numbers and payment details attached.
   2. Write a 3-sentence recommendation: which model tier would you use (Public Web, Private Enterprise, or Local Open-Weights), based on cost, data privacy, and schema compliance (whether it follows the same card format every time)?

   A complete answer

   • Picked one tier and named all three factors: cost, privacy and format
   • Said Public Web is a bad fit while red-light info rides along with the dispatches
   • Mentioned that a person still checks every allergy line before it's used
4. BonusOptional5 min

   Write your resume line

   1. Fill in the blanks: "Audited AI-generated [kind of document, e.g. ingredient cards] for [what you caught, e.g. invented allergy and shipping claims], and rewrote the prompt so the AI flags unconfirmed details instead of guessing."

   A complete answer

   • Both blanks filled with what you actually did