Practical help with AI at work

Ten minutes. No signup required.

Try one useful task in ten minutes

Pick something you can check yourself. Use an AI tool your workplace permits, or use a fictional example in a personal tool. Do not paste customer records, private emails, confidential documents, passwords, or sensitive personal details.

Choose one exercise

1. An office task: turn invented meeting notes into a short action list. Check that each action, owner, and date actually appears in the notes; missing details should say unknown.
2. A customer reply: use a made-up complaint and a made-up return policy. Ask for a polite reply that makes no promises beyond the policy. Check every promise. Do not send it automatically.
3. Learning something new: ask for an explanation of a familiar topic, then three practice questions. Check the explanation against a source you trust. A convincing explanation of a correction does not establish reliability. For an unfamiliar high-stakes subject, have a qualified person check the material before relying on it.

Try this prompt

Help me with this low-stakes task: [task]. Use only the information below. Do not invent names, numbers, deadlines, permissions, or policies. Mark missing information as unknown. Give me a draft and three things I should check before using it. Here is fictional or public information: [input].

Three ready-to-try examples

All the situations below are invented. Add one to the prompt above.

Office: turn notes into actions

Task: make an action list from these meeting notes. The designer will send two mockups by Thursday. The coordinator will check the venue booking; no date was agreed. We also need someone to test the registration form, but no owner was chosen.

Check your result: the designer has a Thursday deadline. The venue check has no deadline. The form test has no owner or deadline. A tool that fills in those gaps without asking has invented facts.

Customer service: draft a bounded reply

Task: draft a reply, but do not send it. Our fictional shop accepts unused products returned within 30 days with a receipt. Customer: I bought this 10 days ago, have not used it, and still have my receipt. Can I return it? No information about shipping charges or refund timing was supplied.

Check your result: the reply can explain the stated return conditions. It must not promise free shipping, invent a return address, or promise when a refund arrives. Those details need checking.

Learning: test an explanation you can check

Task: explain the difference between a rectangle and a square to a beginner. Use these facts: both have four right angles; a square has four equal sides; a rectangle need not have four equal sides. Give three practice questions, then the answers separately.

Check your result: every square fits the rectangle definition here; not every rectangle is a square. Check that its practice answers follow those facts. Fluency alone is not evidence that an explanation is correct.

Keep it to ten minutes

Spend two minutes defining the task, three getting a draft, and five checking it. Stop at ten minutes even if the result is disappointing. Include checking and repair time when deciding whether it helped.

Record one sentence: This helped with ___, but I still needed to ___. A useful failure is a result too. Repeat on several examples before trusting a pattern; one success is not proof of reliability.

If you do the work

Map your work without predicting your fate

Write down five things you actually do in a normal week. Avoid broad labels such as marketing or operations. Use tasks such as summarizing a meeting, checking an invoice, explaining an option, or resolving a complaint.

For each task, answer:

• What result does another person need from me?
• Could an AI prepare a draft with public or invented information?
• What must a person know to spot a bad answer?
• What happens if the answer is wrong?
• Who is allowed to approve or act on it?

Pick one low-risk task to test, one that needs an expert review, and one you will not delegate yet. This is a way to learn about your work, not a score predicting whether your job survives.

For managers

Ask the person doing the work to help define a good result. Do not convert a single successful demonstration into a staffing decision. Ask how newcomers will learn the work if introductory tasks change.

If you are setting up an assistant

Decide what the assistant may do

Before connecting an AI to anything important, finish these sentences:

• The result I want is ___, and a result I would reject is ___.
• The information it may use is ___.
• It may draft ___, but it must not send, spend, delete, or publish without ___.
• The person checking its work is ___, and that person can stop it by ___.
• If it is uncertain, it should ___.
• If something goes wrong, we can undo ___. Actions we cannot undo require ___ before proceeding.

Start with drafts and read-only work. Where a mistake could affect money, someone’s rights, health, safety, or reputation, use qualified review and explicit permission. A person who has no time or authority to say no is not meaningful oversight.

If you run a small business

A small-business workflow brief

Choose one repeated irritation, not a whole-company replacement. Fill in:

Today

We receive ___ and manually turn it into ___.

Frequency

This happens ___ times per week and takes about ___ minutes, including corrections.

A useful draft

An assistant could prepare ___ for a person to check.

Limits

It must not access ___ or promise ___.

A fair test

We will use invented examples first, then an approved small sample. We will compare total time, accuracy, and correction effort with our current process.

Keep or stop

We keep it only if ___ improves without making ___ worse.

Do not assume a low software price means a low total cost. Count setup, checking, training, maintenance, and switching away if it does not work.

If you are weighing the arguments

Read AI policy arguments critically

Use these questions at your own table:

1. Which matters for the claim being made: a strong model, enough hardware and electricity, or the ability to use it reliably in real work?
2. Does a rule delay a useful project, prevent a concrete harm, help people trust it, or do several of those things?
3. Who benefits from new capacity, and who bears local costs?
4. What would count as evidence that our preferred policy is failing?
5. How do people keep a meaningful voice over powerful systems, including systems developed here?

Disagreement is welcome. You do not need to share a particular political view to learn useful things about AI.