Automate a task with AI: a step-by-step plan
Most people use AI one question at a time. The bigger win is handing it a task you do every week, such as sorting emails, filing invoices or writing the Friday update. Here is how to pick the right task, choose how far to automate it, keep a person where it matters, and check it actually saves time.
The rule of thumb: automate one small, boring, checkable task first. Keep a person on anything that is sent, paid or deleted, and test on real past examples before you trust it.
Step 1: pick the right task
Think of one task you do often. Tick each line that is true for it.
- 5 or 6 yes: A good candidate. Worth automating. Start at level 2 or 3 below and test it on real past examples first.
- 3 or 4 yes: Worth a try, with care. Start at level 1 or 2: you stay in the loop on every result, and you learn where it goes wrong.
- 0 to 2 yes: Pick another task. Too much can go wrong, or too little time is saved. Use AI as a helper here, not as an automation.
Step 2: choose how far to automate
One example all the way down: a small online shop that answers 40 customer emails a week. Start at the lightest level that does the job.
1 Do it in a chat, by hand
You paste the input into a chatbot, ask for the result and copy it back. Nothing is set up, and you see every answer.
A customer emails about a late parcel. You paste the email into a chatbot and ask for a short, friendly reply, then edit it and send it yourself.
- Good when
- You do the task now and then, or you are still working out how to describe it.
- Watch out
- You retype the same instructions every time, and results vary with how you ask.
2 Save it as a reusable assistant
You write the instructions once and attach the documents it needs. After that you just paste in the new input.
You set up an assistant with the returns policy, the delivery times and three replies you liked. Now you paste in each email and get a reply in the shop's tone.
- Good when
- The same task comes up every week, with the same rules.
- Watch out
- It still needs you to start it each time, and the documents go out of date if nobody updates them.
3 Build a workflow that runs by itself
Something happens, such as an email arriving, and a fixed chain of steps runs: the AI does one step in the middle, and the result lands where you need it.
Each new email in the shop's inbox is sorted into "Where is my order?", "Return", "Complaint" or "Other". For the first two, a reply is drafted and saved in the inbox. A person reads it and presses Send.
- Good when
- There are many cases, they look alike, and you can describe every step in advance.
- Watch out
- It runs whether you are watching or not. Unusual inputs get the same treatment as normal ones unless you plan for them.
4 Hand the goal to an agent
You give a goal instead of steps. The agent decides what to look up and which tools to use, then reports back or asks before acting.
For a return, the agent looks up the order, checks it is within 30 days, prepares the refund and asks a person to approve it before any money moves.
- Good when
- Cases vary too much to script every step, and a person approves anything that matters.
- Watch out
- The least predictable level: it can take a wrong turn, and hidden instructions in an email can trick it.
Step 3: sketch the workflow
Every automation has the same six parts. Write each one down before you open any tool.
- TriggerWhat starts it.A new email arrives in the shop's inbox.
- InputWhat the AI gets to work with, and nothing more.The email text and the order number, if there is one.
- AI stepOne clear job, written as a prompt."Sort this email into one of four types, then draft a reply using the policy below."
- CheckWhere a person or a rule checks the result.Drafts wait in the inbox; complaints always go to the owner.
- ActionWhat happens to the result.The checked reply is sent, and the email is tagged.
- LogA record of what it did, so you can check later.One row per email in a spreadsheet: type, draft, sent or changed.
Give the AI step one job only. "Sort, then draft" works better as two clear instructions than as one long, vague one.
Step 4: decide who approves what
The more an action matters, the closer a person stays to it. This is called keeping a human in the loop.
| If the automation… | Who decides | Rule |
|---|---|---|
| Reads, sorts or summarises | Runs alone | Let it run. Spot-check a few results each week. |
| Writes a draft for you | Person checks | Let it run. A person reads it before it goes anywhere. |
| Sends something to customers or the public | Person approves | A person approves each one, at least until you trust it. |
| Spends money, changes records or deletes | Person approves | A person always approves. |
| Decides about people: hiring, credit, health, discipline | Person decides | Don't automate the decision. AI may help prepare it; a person decides. |
Is it worth it?
Checking AI's work takes time too. Put in your own numbers; we've started with the shop's emails.
If checking takes almost as long as doing it, automate a different task, or check fewer results once you trust it. Remember to count the time to fix the wrong ones.
Step 5: test it before you trust it
- Collect 20 real past casesInclude the awkward ones: the angry email, the one in another language, the one with no order number.
- Run them all throughPut each one through your prompt or workflow, without fixing anything by hand.
- Sort the results into three pilesRight as it is, needs a small fix, or wrong. Note what kind of mistake each wrong one is.
- Decide where a person checksUse the table above. Start with approval on; you can relax it once the results earn it.
- Check again every monthCustomers, products and the AI model itself all change. A workflow that worked in spring can quietly get worse.
This is a small version of what AI teams call evals. More in how to tell if an AI system works.
What not to automate
- Anything you can't checkIf nobody would notice a wrong answer, you will only find out when it does damage.
- Final decisions about peopleHiring, credit, health and discipline need a person who can explain the decision. In the EU, AI used for most of these is "high risk" with strict rules. EU AI Act →
- Messages that need careBad news, condolences, a complaint that has already gone wrong. People can tell, and it costs trust.
- More access than the task needsConnect only the apps and folders it uses, never share passwords or card details, and keep confidential data in tools your organisation has approved. AI and your privacy →
- Workflows that read outside text and then actAn email or web page can hide instructions meant for the AI. If it reads them and can also send, pay or delete, keep a person in between. Prompt injection →
Three recipes to start with
Each one works as a saved assistant (level 2) or as a workflow (level 3).
Meeting notes into an action list
- Trigger
- A meeting transcript is saved.
- AI step
- Pull out decisions, who owns each action, and deadlines. Mark anything unclear as "to confirm".
- Action
- Send the list to you as a draft to share with the team.
- Check
- You check names and dates before sharing: transcripts often mishear them.
Invoices from email into a spreadsheet
- Trigger
- An email with an invoice attached arrives.
- AI step
- Read the supplier, invoice number, date, amount and due date.
- Action
- Add one row to a spreadsheet, marked "to check".
- Check
- A person checks totals before anything is paid. The workflow never pays.
A weekly update, drafted for you
- Trigger
- Every Friday at 3 pm.
- AI step
- Read this week's rows from the team's task or sales sheet and write a five-line summary.
- Action
- Save it as a draft email to your manager.
- Check
- You read it, add what the numbers don't show, and send it yourself.
How we put this guide together
The four levels and the six parts of a workflow are how automation tools such as Zapier, Make, n8n and Microsoft Power Automate, and the AI companies’ own assistant and agent features, describe what they do. The scorecard, approval rules and testing steps are our own advice, based on the other guides on this site. The online shop is an illustrative example, not a real business, and the calculator only does the sums you give it. Tool names change often, so we don’t compare features or prices; check each tool’s own site. We review this guide every quarter. Spotted something out of date? Tell us.
Keep exploring
For the tools themselves, see AI agents explained, which lists ready-made agents and the builders mentioned here. To choose between buying a tool and building your own, see build or buy AI. For the tools that suit your job, see AI tools by job.