What is AI actually good at for a normal person?

First drafts (emails, posts, proposals), summaries (calls, long threads, documents) and sorting messy information (reviews, notes, receipts). If a task is written, repeats every week and you can check the result quickly, it is a good first candidate. Check every draft before it goes out.

What should I type into ChatGPT or Claude first?

Use the prompt in the worked example below, with your five tasks. It answers with a ranked list you can act on today.

What is AI still bad at?

Facts it cannot check, exact numbers, anything that needs your judgment about a person, and jobs with no clear finish line. Use it to draft and sort, and keep the final call yourself.

What Can AI Do for Me? Start With What Ate Your Week

Q: What can AI do for me?

AI is most useful on work you already do: tasks that repeat, are mostly writing or reading, and that you can check in seconds. Don't ask it what it can do. List five tasks that ate your week, paste them into ChatGPT or Claude, ask it to rank them, and try the top one on real work.

If you run a business or a team, you have a week full of tasks an AI could draft. Give the AI that week and it can rank them. Asking “what can you do?” only gets a list that fits anyone.

The five-task method

Write down five things you did this week that you hated or that took too long, one line each, with how long each took and how often you do it
Open ChatGPT or Claude
Paste the prompt below with your list and ask which it could draft and what it needs from you
Try the top-ranked task on this week’s real work

Worked example: a bookkeeper’s week

The prompt to copy

Here are five tasks from my week as a freelance bookkeeper, with the hours each took (sample numbers). Rank them by how well an AI assistant could do a first draft. For each, say what you would need from me, and what I must check myself. Do not make up details.

Chasing clients for missing receipts (3 h)

Writing the monthly summary email to each client (4 h)

Categorising bank transactions (5 h)

Answering the same tax-deadline questions (2 h)

Deciding which clients to drop (1 h)

Sample output (written by us to show the shape; your answer will differ)

Rank	Task	AI role	You keep
1	Monthly summary email	Draft from your figures and last month’s email	Check every number against the books
2	Receipt chasers	Draft a polite 3-step reminder sequence	Decide who gets which reminder
3	Tax-deadline questions	Draft a reusable answer sheet	Confirm each deadline with the tax authority
4	Categorising transactions	Suggest categories	Review each one; errors cost money
5	Dropping clients	Nothing useful	The whole decision

Read the ranking by one test: a draft you can check in under a minute goes first. Task 4 sits low because wrong categories are costly and hard to spot at volume.

What makes a good first task

Good first task	Bad first task
Repeats every week	Happens once a year
Mostly writing or reading	Mostly judgment about people
You can check the result in seconds	You can’t tell if it’s right
You dread it	You enjoy it

Three or four on the left means start there.

Where it goes wrong

The AI states something false in a confident tone, usually a number, date or name you supplied incompletely. Spot it by checking every figure in the draft against the source before sending. A ranking that puts a hard judgment call on top is the same warning: you gave it too little detail, so add hours and context and ask again.

Don’t paste client or customer data into a consumer chat. Anthropic says it uses consumer chats for model training only if you allow it, or if a chat is flagged for safety review (Anthropic, checked Oct 31, 2026). OpenAI’s data-control page blocked our check, so confirm the setting on OpenAI’s help page. Business plans differ again. Remove names first, as covered in what not to tell ChatGPT.

Next step

Put your five tasks into the AI ROI calculator. It ranks them by hours saved and writes a ready prompt for your top three, so you start with the task worth the most.