Start with something you can verify
A useful first AI task has a clear input, a small output and a check you already know how to perform. Turning a short set of notes into an action list is a better learning exercise than asking a tool to decide your team’s strategy.
Before trying this at work, check which tools and data your employer permits. Use public or invented information for practice. Removing a customer’s name does not necessarily make a record safe to share: the remaining details can still identify them or reveal confidential work. The NCSC explains why unapproved AI tools create data risks.
- Good starting point: a short, repeatable task with a result you can check against a source.
- Too open-ended: a recommendation where you cannot tell whether the reasoning or evidence is sound.
- Stop condition: you need data you cannot share, permission is unclear, or no qualified person can review the result.
A tidy draft can still invent a deadline
AI-generated text can sound convincing while containing unsupported details. NIST calls this risk confabulation. Practice checking the evidence before you judge the polish.
The source: fictional meeting notes
Meeting date: 6 October. All people and work below are invented.
- Maya will send the revised supplier brief by 9 October.
- Louis will compare two public pricing pages. No deadline was agreed.
- The owner of the customer call is undecided. Assign someone at the next meeting.
Would you send this action list?
This practice draft contains deliberate mistakes. Compare every owner and date with the notes above.
Practice draft. It has not been verified.
Send the revised supplier brief
- Owner
- Maya
- Deadline
- 9 October
Compare two public pricing pages
- Owner
- Louis
- Deadline
- 9 October
Book the customer call
- Owner
- Louis
- Deadline
- 10 October
Read the answer key
- Both owner and deadline appear in the notes.
- The draft invented a deadline. Keep it unassigned until it is confirmed.
- Neither the owner nor the deadline was decided. The next step is to clarify them.
The important correction is sometimes “not agreed.” Filling every cell makes a document look complete; leaving an uncertainty visible makes it easier to resolve. The example is written by us to teach that distinction. It is not output from a live model or a measured error rate.
A brief you can try with the fictional notes
Turn these notes into an action list with task, owner and deadline. Use only details stated in the notes. Write “not assigned” or “not agreed” where information is missing. List unresolved decisions separately. Do not infer a deadline from the meeting date.
Clear instructions help define the task. They do not guarantee the output will follow them. Check each row against the original notes, even if the tool says it verified its answer.
Count the checking time, too
For your own trial, make one version using your normal method and another using an approved tool. Keep the input and quality standard comparable. Record preparation, drafting and review time separately, along with any errors that survive your first check.
| Method | Draft | Check | Total |
|---|---|---|---|
| Usual method | 8 min | 2 min | 10 min |
| With AI | 3 min | 8 min | 11 min |
Here the faster draft creates more checking work, so it takes longer overall. A trial that reveals this is useful: you can change the brief, choose a different task or stop. One good result is also too little to justify handing a whole workflow over to automation.
If you work with numbers, try the checked examples for financial analysis or data analysis. They add arithmetic, duplicate records and metric definitions to the review.
Leave with one experiment to run
Make a small commitment you can review
Choose a task, a time budget and a quality check. The free worksheet keeps your entries in the current page; copy or print them before leaving. Use public or invented information.
Unsure where to start? Take the free assessment to reflect on the work you do. Or use our 30-day career preparation guide to turn the experiment into a learning plan.
Sources & limits
These sources support the ideas about AI, work and responsible use. The exercises are our teaching examples. They have not been validated as a way to prevent job loss or improve earnings.
- NCSC: The hidden risks of shadow AI
7 September 2026
Guidance on unapproved workplace tools and loss of control over company data. Your employer’s own permission and data rules still apply.
- NIST: Generative Artificial Intelligence Profile (AI 600-1)
26 July 2024
Describes risks including plausible false output and data privacy. It does not certify the exercise below or tell you how accurate a particular tool will be.