How to automate tasks with AI in a small business

To automate a task with AI, pick one job you repeat every week, write its rules in plain words, and have the AI produce a first draft that a person checks before anything is sent. Start with the job that comes up most often and is quick to check. Payments, complaints and legal decisions stay with a person.

Further down there's a prompt that picks your first task, with the results from 18 test runs on 9 November 2026, and a guide for each of eight common admin jobs.

What you need

• An AI chat you already use. We ran the prompt on this page in Claude. It's plain text, so it should paste into ChatGPT or Gemini too, but we haven't run it there.
• A list of the jobs you or your staff repeat: what each one is, how often it comes up, how long it takes and who does it.
• Ten old examples of the job you end up choosing, with names and phone numbers removed.
• About 30 minutes for the list and the prompt. Building the first job takes longer, and each guide below has its own steps.

What does automating a task with AI mean in a small business?

In most small businesses it means the AI writes the first draft of a repeated job and a person checks it. The AI may cut the time you spend drafting. Preparing what it reads, correcting its drafts and checking the result still take time.

It comes in three sizes.

1. A saved prompt you run by hand. You paste in today's messages or documents, the AI drafts, and you check and send. The eight job guides further down all work this way, and you don't install anything.
2. A scheduled job. The same kind of prompt runs on a timer and leaves its result where you'll see it. Our daily AI news digest guide shows one. It's how this site's news page was first produced: GPT did the research every morning, and a script checked every link before the day's file was written.
3. AI connected to your tools, so it reads your inbox, files or calendar directly. Connect Claude to Gmail, Drive and Calendar safely covers the read-only tests and approval rules to set up first.

Start with the first size. A prompt that gives wrong drafts by hand will give the same wrong drafts on a timer, only with nobody watching. Move up once the saved prompt has given you drafts you barely edit for a few weeks.

Which tasks should you automate first?

Look for a job that passes four tests.

• It comes up often. A job you do 25 times a day pays back sooner than one you do once a month.
• You can write its rules down. If you could explain the job to a new staff member on one page, you can explain it to an AI.
• You can check the result in under two minutes against a price list, a diary or the original document.
• A wrong draft costs little, because a person reads it before it goes anywhere.

Some jobs fail the last test however often they come up: paying suppliers, answering complaints, refunds, legal or safety decisions, hiring and firing. AI can help you prepare for these, and a person still decides and sends. The five things AI should never do without you turns that into a one-page rules file.

A prompt that picks your first task

Copy this into a new chat. Replace the two bracketed parts at the end with your own business and task list.

You are helping a small business owner decide which repeated task to hand to AI first. Use only what I write below. Do not invent tasks, numbers or tools.

For each task in my list, give:
1. How often it happens and how long it takes each time, from my notes. Write "not stated" if I didn't say.
2. Whether the rules for doing it can be written down in plain words (yes, partly or no), with one line on why.
3. What could go wrong if the AI gets it wrong, and who would notice.
4. How I would check the result in under two minutes.
5. A verdict: DRAFT (AI writes a first draft, a person checks and sends), HELP (AI helps prepare, a person does the task) or KEEP (stays with a person).

Rules:
- AI never sends, pays, deletes or promises anything. Every DRAFT task ends with a person approving.
- Anything where money leaves the business, a complaint, a refund, a legal or safety decision, or hiring and firing is KEEP or HELP, never DRAFT.
- If my notes are too thin to judge a task, write NEED MORE and ask me one question.
- Work out hours per week only from my numbers. If a number you need is missing, such as how many days a week we work, give the figure my numbers do support (hours per day, for example) and ask me for the missing one. Do not fill it in yourself, even under an ASSUMPTION label. Show the working and tell me to check it.
- If you assume anything I didn't write, start that sentence with ASSUMPTION.
- If you suggest a number that isn't in my notes, such as how many old examples to test, say it is your suggestion.

Then pick ONE task with a DRAFT verdict to start with this week: the one that happens most often, has rules I can write down, and is easy to check. Say why in two sentences. List what I need to prepare (three items at most). Name the first test I should run on old examples, with names and phone numbers removed, before I use it on real work.

If no task gets a DRAFT verdict, do not pick one. Say so, and tell me what you would need to know before any task could be handed over.

My business: [one line]

My repeated tasks:
[one task per line: what it is, how often, how long, who does it]

What happened when we ran it

We ran the prompt on 9 November 2026 in fresh Claude sessions, using Claude Sonnet 5.5 through Claude Code with tools and our own settings switched off. The business was made up: a six-person air-conditioning service company in Kuala Lumpur. This is the task list we pasted in.

My business: Cool Breeze Aircon Services, a made-up six-person air-conditioning service company in Kuala Lumpur. Two office staff, four technicians.

My repeated tasks:
1. Replying to WhatsApp enquiries about servicing prices and available slots. About 25 a day, 3 minutes each. Office staff.
2. Writing quotes for installation jobs from the customer's message and our price list. About 8 a week, 20 minutes each. Me.
3. Following up on quotes that got no reply. About 10 a week, 5 minutes each. Me, when I remember.
4. Paying supplier invoices by bank transfer. About 12 a month, 10 minutes each. Me.
5. Replying to complaints about a job. About 2 a week, 30 minutes each. Me.
6. Weekly summary of jobs done and money collected, for my business partner. Once a week, 90 minutes. Office staff.
7. Stock.
8. Sorting technicians' fuel and parts receipts for the accountant. Once a month, 3 hours. Office staff.

Four of those lines are traps: task 4 moves money out of the business, task 5 is a complaint, task 7 is a single word, and task 1 gives a daily count with no working days, so an honest answer can't turn it into hours per week.

The prompt above is the fourth version. The first one didn't have the last two rules or the closing paragraph about lists with no DRAFT task, and its rule about hours was a single line: work out hours per week only from my numbers, show the working and tell me to check it. We ran it three times.

• All three runs picked the WhatsApp enquiries as the first task, with the AI drafting and office staff sending.
• All three kept supplier payments with the owner. Two kept complaints with the owner too. The third let the AI summarise the complaint while the owner writes the reply.
• All three wrote NEED MORE for "Stock" and asked what the owner does with stock.
• None of them turned 25 enquiries a day into a weekly figure. Each gave 1.25 hours a day and told us to multiply by our own working days.
• We checked the other sums by hand, for example 8 quotes at 20 minutes is 160 minutes, or about 2.7 hours a week. They were right.

Two things went wrong. Each run suggested how many old enquiries to test on (20, 30 and 20), and only one of the three said the number was its own idea. The weekly summary got HELP in one run and DRAFT in the other two. Our notes never said where the job and payment figures are kept, and one run wrote that it was assuming they sit in one place.

So we added the two rules about assumptions and suggested numbers and ran the second version three times.

• All three said the test size, 20 old enquiries, was their own suggestion.
• All three stopped guessing about the weekly summary. It came back NEED MORE, with a question about where the figures are recorded.
• All three still picked the WhatsApp enquiries.
• The complaint verdict still moved: HELP in two runs, KEEP in one. One run also wanted more detail on the receipts before giving a verdict.

In the second version a new problem appeared: all three runs assumed the working days. Each turned the 25 enquiries a day into a weekly figure, 6.25 hours at five days, and two runs added 7.5 hours at six. Every figure carried the ASSUMPTION label. The prompt, though, now said to use only our numbers and, on the next line, how to label a number it had assumed.

The third version says what to do when a number is missing: give the figure the notes support, ask for the missing one, and leave it blank until the owner answers. We ran it three more times.

• None of the three gave a weekly figure for the enquiries. Each gave 1.25 hours a day and asked us for the working days.
• All three picked the WhatsApp enquiries again, and all three said the 20 old enquiries were their own suggestion.
• The weekly summary and "Stock" came back NEED MORE in all three.
• Paying suppliers stayed with a person in all three. Two runs gave KEEP. The third gave HELP, with the AI laying out the invoice details for the owner to check before paying.
• Complaints got HELP in two runs and KEEP in one. Receipts got HELP in two and DRAFT in one.

The aircon list couldn't show one more gap, because it always had a task that was safe to draft. Every version so far told the AI to pick one task and never said what to do if none was safe to draft. So we wrote a second made-up list for a two-person bookkeeping firm: paying salaries and bills, answering a fee dispute, refunding a deposit, hiring temporary help, and one line that only said "Filing". The rules bar four of those from DRAFT, and the fifth is too thin to judge. We ran the third version on it three times.

• No task got a DRAFT verdict in any run. "Filing" came back NEED MORE each time.
• All three picked a task anyway. Two chose the salary and bill payments as a HELP task, and one chose the deposit refunds.

The fourth version picks only from DRAFT tasks and says what to do when there are none. We ran it three times on each list.

• On the bookkeeping list none of the three picked a task. Each said no task had a DRAFT verdict and asked what the filing involves, since it was the only job that might qualify.
• On the aircon list all three picked the WhatsApp enquiries, kept supplier payments with the owner, asked for the working days and said the 20 old enquiries were their own suggestion.
• Complaints on the aircon list got HELP in two runs and KEEP in one. Receipts got HELP in all three.

The 18 runs took 27 to 42 seconds each.

This test has limits. No real business used the result, so it gives no time-saved figure, and the hours come from numbers we made up. Eighteen runs of two task lists in one model is a small sample. On the aircon list the verdict moved on four tasks across the twelve runs: complaints, the weekly summary, receipts and, once, supplier payments. On the bookkeeping list it moved on the payments and the fee dispute. So run the prompt twice on your own list. If either answer says NEED MORE for a task, answer its question before you choose that task. For any other difference, take the more cautious verdict.

How do you build the first automation?

The eight guides below share one method.

1. Write the rules in your own words: what a good result looks like, what the AI must never do, and when it should stop and ask you.
2. Gather what the job reads from: the receipts, the invoices or the files your report is built from. For quotes and replies about prices, that is your current price list, and Put your price list in a shape AI can read correctly shows the layout.
3. Collect ten old examples and take out names, phone numbers and addresses. Before you paste anything into AI: a two-minute safety check lists what else to remove.
4. Run the prompt on the old examples and compare each draft with what you really sent.
5. When a draft is wrong, change the rule that allowed it and run all ten again. If you only fix the draft, the same mistake comes back the next time.
6. Save a dated copy of the version that works.
7. Use it on real work with a person reading every draft. Check the AI's answer before a customer sees it gives that person six questions to ask.

Which guide fits which job?

• New enquiries piling up: The Lead Triage Desk sorts them into Hot, Warm, Later and Unclear, drafts replies for the safe ones and flags the risky ones.
• WhatsApp questions about prices and hours: Draft WhatsApp replies in BM or English that you check and send works from a notes file of your real prices. We tested it on a made-up shop, 12 runs in fresh sessions, on 4 October 2026.
• Quotes: Turn a customer message into a draft quote builds line items only from your price list and leaves the totals to a spreadsheet.
• Quotes nobody answered: Follow-ups that stop when they should plans follow-ups with a clear end and stop rules.
• Supplier invoices and service agreements: Invoice and agreement intake logs them and works out renewal deadlines. Its prompts were tested in Claude on made-up paperwork on 23 September 2026.
• Receipts for the accountant: Sort a pile of receipts into a clean expense table writes UNREADABLE where it can't read a figure, and your spreadsheet does the totals.
• The Monday report: A weekly business report you check instead of write fills the same short report from your files and traces every number to a file and column.
• Explaining a job to staff: Turn a rambling voice note into a step-by-step SOP turns a recording of you talking through the job into numbered steps, plus the questions you left open.

The business workflows page puts six of these in the order an owner would set them up, with the safety check, the price list and the final check around them. The guides are starter versions of the Desks taught in the Academy.

When should you move past a saved prompt?

Move up one step at a time, and only for a job that already works by hand.

If you paste the same price list and rules into every chat, put them in a Project so each chat starts with them. Set up a Claude Project as your business brain covers Claude, and the same setup works in a ChatGPT Project or a Gemini Gem. Once a job has been stable for a few weeks, you can package it as a Skill with Write your first Claude Skill for quotes or enquiry replies. Connect your inbox or files last, and start with read-only access.

Questions people ask

What is AI automation? In a small business, it's AI doing a repeated job by a written set of rules. Usually that means the first draft, with a person approving the result.

Do I need to know how to code? For the eight job guides, no. Each one uses prompts you paste into a chat, your own files and checks you do yourself. The daily news digest linked above does use a script. The Gmail, Drive and Calendar guide is set up through menus in the app.

Which AI tool should I use? The one you already have. Seven of the eight job guides are written for Claude, ChatGPT or Gemini. The invoice guide was tested only in Claude.

How much time will it save? We don't have a measured figure from a real business. Time a few examples both ways, and on the AI side count preparing the input, correcting the draft and checking it. Subtract the time with AI from the time without it, then multiply the difference by how often the job comes up each week.

Is it safe to paste customer messages into AI? Take out names, phone numbers, addresses and bank details first, and keep some jobs out of AI altogether. The two-minute safety check has the list.

Quick checklist

• List your repeated jobs with how often, how long and who.
• Run the prompt twice. Answer any NEED MORE question, and keep the more cautious verdict where the two runs differ.
• Pick one job that is frequent, has rules you can write down and is quick to check.
• Keep payments, complaints, refunds, legal and safety decisions with a person.
• Test on ten old examples with names and numbers removed.
• Fix the rule, then rerun every example.
• Save a dated copy of the version that works.
• A person reads every draft before it goes out.