What you’ll be able to do

Right, so this is about the boring stuff. The weekly report you copy into the same spreadsheet every Friday. The form you fill in with the same twelve fields, different customer. The email you triage every morning before you’ve even had a coffee. All of that can be handed off, at least partly, to an AI while you keep your hand on the wheel.

You won’t build a robot that runs your business. You’ll build a small, dependable workflow that does one repetitive job, and you watch it do that job until you trust it. That’s the whole ambition here. Not autonomy, just less typing.

Before you start

Not every task is a good candidate. The ones that work well are the ones with clear steps, a predictable shape, and low stakes if something goes slightly wrong. Filling in a supplier form with the same details each time? Good. Sorting your inbox into folders based on sender? Fine. Deciding which client gets a discount and how much? That’s a judgement call, and you want a human on that, always.

Before you touch a tool, write the task down as if you’re training a temp on their first day. Literally, steps: open this, copy that, paste it there, check this box. If you can’t write it in plain steps, the AI can’t do it reliably either, and it’ll just make things up when it gets stuck.

Also worth knowing upfront: agents that operate a browser or a computer, like ChatGPT’s agent mode or Anthropic’s computer-use capability, are watching-required tools. They click buttons and fill boxes on your behalf, but you’re meant to be sat there, at least at first, checking each step. Perplexity’s Comet browser does something similar for web tasks. None of these are “set and forget”. Anyone who tells you otherwise hasn’t watched one of these tools misclick yet. It’s recoverable, but it’s not invisible.

Get set up

There’s no single right tool here, it depends on what you’re automating and where it lives.

If your task is about connecting apps, moving data between them, or triggering something when an email arrives, look at Zapier, Make, or n8n. All three now let you drop an AI step into the middle of a workflow, so a step can summarise a document, classify an email as urgent or not, or draft a reply, sitting between two boring, deterministic steps that just move data around. That combination, rules for the predictable bits and AI for the bit that needs judgement, is the whole trick.

If you’re on a Mac or iPhone, Apple Shortcuts is built in and free, and it can chain simple actions together without needing any of the above. On Windows, Microsoft Power Automate does a similar job and plugs into Office apps directly. Neither needs you to sign up for anything new, which is a nice change.

If the task involves clicking through a website, like filling in a government form or navigating a portal that doesn’t have an API, that’s where the browser-operating agents come in. ChatGPT’s agent mode and Anthropic’s computer use are built for exactly this: they see a screen, they click, they type, and critically, you watch them do it.

Setup, practically:

  1. Pick one task. Just one. The weekly report, or the form, not both.
  2. Write it as numbered steps, the temp-on-day-one version.
  3. Decide which steps are rules (always do X) and which need judgement (decide whether X applies). Rules go to automation software. Judgement goes to the AI step.
  4. Connect your accounts in whichever tool you’ve picked, following their own connection flow, don’t paste passwords into a chat window.
  5. Build the smallest version that could possibly work. Test it on one example before you let it near real data.

That’s it. Don’t build the whole pipeline in one sitting. Build the ugly version first.

Try it yourself

Here are three real prompts, each doing something different. Adjust the details to match your own task, obviously, these are templates, not incantations.

For a form-filling job, something you’d use inside an agent that can see a browser:

I need you to fill in the supplier registration form at [URL].
Use this information:
- Company name: [X]
- Contact email: [X]
- VAT number: [X]
- Address: [X]

Fill in each field, then stop before clicking submit.
Show me a summary of what you entered so I can check it,
and wait for me to say "go" before you submit anything.

For a weekly chore, say a report you pull together from a few sources every Friday:

Every Friday I need a short summary for my manager covering:
1. Tickets closed this week (I'll paste the list below)
2. Any tickets still open after 5 days
3. One line on overall workload compared to last week

Here's this week's data: [paste]

Write it as three short paragraphs, plain language,
no bullet points, ready to paste into an email.

For copy-paste work, moving the same kind of information from one format into another:

I'm going to paste in a block of customer feedback.
For each entry, pull out:
- the customer's main complaint (one sentence)
- whether it's about price, delivery, or product quality
- whether it needs a reply today or can wait

Give me the answer as a simple table, one row per entry.
Don't add commentary, just the table.

Notice none of these ask the AI to decide anything irreversible. They ask it to draft, summarise, sort, or prepare, and then stop and wait. That pause is doing a lot of work. Keep it in.

Check the result

Don’t trust the first output because it looks tidy. Tidy and correct are different things, and AI is very good at tidy.

For form-filling, check every field against the source data, one by one, before you let anything submit. Yes, every field, even the boring ones. Typos creep in on the fields you stopped paying attention to.

For the weekly report, read it as if you’re the manager receiving it. Does the number of “tickets closed” actually match what you’d expect? Pull one number out and check it against the raw data yourself. If it’s wrong once, it can be wrong elsewhere too, and you won’t spot a pattern from one check.

For the copy-paste table, spot-check three or four rows against the original text. Look specifically for entries the AI might have merged, split wrongly, or skipped altogether when the wording was awkward. That’s the failure mode to watch for, not wild invention, just quiet omission.

A good general test: run the same task twice with the same input, a day apart, and compare. If the two outputs disagree on something factual, like a total or a category, you’ve found a weak spot before it cost you anything.

If it doesn’t work

Agents click the wrong thing sometimes. Not often, but sometimes, and usually at the worst possible moment, like the one field where it matters. That’s not a bug you can fully engineer away yet, it’s a reason to keep a human approval step before anything is sent, paid, or deleted. Always. Even after weeks of it working fine.

Never give an automation your banking details or payment credentials to act on directly. If a workflow needs to touch money, that’s a step you do yourself, or a step with hard limits and explicit confirmation built in, not a “just handle it” instruction.

Keep any secrets, API keys, passwords, tokens, in the tool’s own secret store, whether that’s Zapier’s, Make’s, n8n’s, or whatever credential manager your platform offers. Never type them into a prompt. A prompt can end up logged, shared, or pasted somewhere you didn’t intend, and secrets typed in plain text don’t belong there.

If a workflow keeps failing at the same step, that’s usually a sign the step needs a rule, not more AI. Judgement-heavy AI steps are for the bits that genuinely vary. Everything predictable should be handled by the boring, deterministic part of your automation, because rules don’t hallucinate.

Keep exploring

Once this one chore is running reliably, the same approach, steps, rules, judgement, approval, applies well beyond forms and reports. If your inbox is the actual bottleneck, there’s a guide on this site called “How to get AI to handle my emails” that goes deeper on that specific chore. If it’s your files and folders that are a mess, see “How to get AI to sort and organise my files”. And if part of your repetitive task involves scripts or small bits of code to glue tools together, “How to get AI to write code” covers that ground properly.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.