The five duties

Whatever tools your organisation uses, and whatever your job, using AI well comes down to five duties. They're the same for the chief executive and for someone who started on Monday. This lesson takes them one at a time, with the most time on the one people find hardest: knowing where the tools go wrong.

1. Use approved tools

An approved tool is one your organisation has checked and signed up to: how it handles your data, who controls its settings, what it costs. It might be a separate AI assistant, or an AI feature inside software you already use, like your email or documents. A free tool on a personal account hasn't been through any of that, however good it is.

If there's a tool you'd like to use that isn't on the list, ask. Your organisation would much rather hear about it than have it used quietly.

2. Follow the data rules

Not all information is equal. Most organisations sort it into a few classes, and each class says which tools may hold it. A common starting point has four: public (already published), internal (plans and notes with no personal details), confidential (client work, contracts, personal details) and regulated (health records, card data, anything a specific law covers).

The more sensitive the class, the fewer tools it may go into. Your organisation's version is called the data map. It may make a class stricter than the one here, never looser. When you're not sure, treat it as the stricter class, and ask. Try sorting these.

3. Meet the output standard

Anything AI helped you make should be labelled, owned, short and checked before it goes to someone else. The output standard is the whole of the next lesson.

4. Know the tools' limits

AI tools are useful because they're fluent. They write quickly, summarise long documents and turn rough notes into clean drafts. The same fluency makes their mistakes easy to miss, because a wrong answer reads just as well as a right one. Five limits matter most at work.

• Confident wrong answers. A tool can get a sum, a date or a fact wrong in the same assured tone as everything else.
• Made-up sources. Asked for evidence, a tool can produce references, quotes and case names that don't exist. In 2023 a New York court fined two lawyers and their firm for filing a brief that cited cases a chatbot had invented (Mata v. Avianca).
• Out-of-date knowledge. A tool learns from material up to a cutoff date. Prices, rules and products change after that, and it may not say so.
• Instructions hidden in documents. A file, email or web page can carry text aimed at the tool rather than at you, and the tool may follow it. Security specialists rank this, called prompt injection, as the top risk for AI applications (OWASP Top 10 for LLM Applications, 2025).
• Bias when it ranks or judges people. A tool can learn unfair patterns from the data it was trained on, and its reasons can sound perfectly neutral.

None of this means don't use AI. It means knowing what to check. Try it: here are five answers from an AI tool to ordinary requests. In each one, find the part you wouldn't trust, then name the limit behind it.

5. Report problems without fear

When AI gets something wrong, or you see it about to cause harm, stop it if you safely can and tell someone. That includes your own mistakes. The promises from lesson 1 protect you when you do: a mistake you report is handled as coaching, not discipline. Lesson 5 covers exactly who to tell and what happens next.

Now go through the five duties and tick each one when you're clear what it asks of you. They go on your completion record.

Your duties and approved tools are saved to your completion record. Next: the output standard, the four rules for anything AI helped you make.