How to automate repetitive tasks
with AI
Identifying repetitive tasks worth automating, choosing the right tool and getting your first automation into real use without a six-month IT project.
10 minute read
"Automating with AI" sounds like a big project. It almost never is. The automations that actually work in SMEs are small and targeted, and they come from observing what wastes time every week. This guide is the path from daily drudgery to your first live automation — without going through a full-blown IT undertaking.
What makes a task
a good candidate
Not everything repetitive should be automated, and not everything tedious is worth the effort. A task is a good candidate when it has all of these traits together:
- It happens frequently — every day or every week, not twice a year.
- It follows a recognisable logic — you could explain it to a new colleague in a few minutes.
- It works on text or data — emails, documents, PDFs, spreadsheets; not physical work or sensitive relationships.
- The result is verifiable at a glance — whoever checks it can tell immediately whether it's right.
A concrete method for finding them: for one week, ask your office staff to jot down the tasks they repeat most often and find tedious. No tools needed — a list on paper will do. By the end of the week you'll have ten candidates without having thought about it in the abstract.
Typical (and low-risk) examples to start with
- Sorting and categorising incoming emails by priority or topic
- Extracting data from supplier delivery notes and purchase orders into a spreadsheet
- Generating draft replies to recurring requests
- Summarising long documents before a meeting
- Reformatting content for publication across different channels
- Translating business correspondence into your working languages
A back-of-the-envelope calculation
in two minutes
Before touching any tool, do a quick calculation. You need a number, not a business plan:
(times per month) × (minutes each time) = minutes saved per monthA task that occurs 200 times a month and takes 4 minutes each time amounts to over 13 hours a month: it's worth automating. One that occurs 5 times a month and takes 10 minutes each time is less than an hour: leave it alone, or only automate it if it's trivial. Above a few hours a month, it usually pays off; below that, rarely.
Keep the real cost in mind: it's not the AI (requests cost fractions of a penny on affordable models) — it's the time to set up, validate and maintain the automation. An automation that saves you an hour a month but takes three days of work to build won't pay for itself for years.
Four levels,
in order of effort
There is no single "right tool": there is the right one for your level and for that task. In order of increasing effort:
| Tool | When it makes sense | Effort |
|---|---|---|
| Chat (web/app interface) | Occasional tasks or to validate an idea: paste the input, copy the output | Minimal |
| No-code platforms (Make, Zapier, n8n) | Connecting email, spreadsheets and business software with an AI step in the middle, without writing code | Low |
| Scripts + API | Custom logic, high volumes, integration with internal systems | Medium |
| Bespoke development | Only when the use case is established and business-critical | High |
The golden rule: move down a level only when the one above is no longer enough. Many useful automations live happily on a no-code platform for years. Starting with bespoke development for a first experiment is the classic way to turn a two-hour idea into a two-month project.
Step by step,
with a real example
Let's take a concrete example — extracting data from supplier delivery notes into a spreadsheet — and follow it from start to finish. The same pattern applies to almost everything.
- Do the task manually, once, using chat. Take a real delivery note, paste it (or upload it) and ask: "Extract the number, date, supplier and line items in table format." If the result is good, you've proved the case works. If it isn't, automation won't fix it.
- Write the instruction (the prompt) carefully. Specify exactly which fields you want, in what format, and what to do in ambiguous cases ("if the date is missing, write 'missing'"). A precise instruction is half the work.
- Test it on 10 different examples. Not one carefully chosen example — ten real ones, including the messy ones. That's where you discover edge cases.
- Choose the minimum tool that handles the volume. For a few delivery notes a week, even a spreadsheet with a script is enough. For hundreds a day, you need a proper automation via the API.
- Put it into use with a human in the loop. For the first few weeks, a person reviews every output before it enters a system. Note how often it was already correct.
- Reduce oversight only when the numbers support it. When the "right first time" rate is consistently high, ease supervision on straightforward cases and keep it on rare or unusual ones.
Putting oversight
where it's truly needed
"Human oversight" doesn't mean checking everything forever: it means placing control where it matters. Three practical ways to do this:
- Approval before sending. The AI prepares, the person confirms with a click. Suitable for emails and external communications.
- Spot-check review. Once quality is high, verify a sample of outputs rather than all of them. Suitable for high-volume data extraction.
- Thresholds and exceptions. The automation handles clear-cut cases on its own and queues only uncertain ones for a person (unusual amounts, missing fields, rare cases).
The mistakes that
derail automations
- Starting too big. "Let's automate the entire purchasing department" is not a project — it's ten projects. Pick one.
- Skipping the manual test. Building the automation before verifying that the model can handle the task is the quickest way to waste days.
- Choosing the most powerful tool. Bespoke development for a first experiment is almost always over-engineered. No-code until it stops being enough.
- Measuring nothing. Without a "right first time" rate, you'll never know whether you can trust it, and the automation stays an eternal experiment.
- Forgetting maintenance. Formats change, suppliers change document templates. An automation needs an owner who keeps it running.
In summary
Find a repetitive task that works on text or data and can be verified at a glance; calculate the minutes saved; test it manually before automating it; choose the minimum tool that handles the volume; keep a person in the loop until the numbers give you confidence. The first useful automation can be live in days, not months — as long as you keep it small.
Frequently asked
questions
Where do I start if I don't have a clear use case?
For one week, ask your office staff to note down the tasks they repeat most often and find tedious. No tool needed — a list on paper is enough. By the end of the week you'll have 10 candidates without having thought about it in the abstract, and you can run them through the five-question test.
Which no-code tools do you recommend?
For those starting from scratch: Make (more powerful, moderate learning curve) and Zapier (simpler, slightly more expensive). For those who want to self-host and have technical skills: n8n. All three have ready-made blocks for calling AI models without writing code.
When is it worth moving from no-code to a custom script?
When volumes grow (hundreds of runs per day) or when the use case requires logic that no-code doesn't handle well (e.g. integrations with internal systems, complex conditional checks, critical performance requirements). As long as no-code holds up, stay there.
How much time does it take to maintain an automation once it's in production?
Typically a few hours to a day per month, mainly to handle the edge cases that surface over time (a supplier who changes their invoice format, a new field to extract). An automation without an owner who keeps it alive will break silently.
How do I measure whether the automation is working?
Two numbers are enough: the "right first time" rate (on how many human-validated outputs the AI was correct without corrections) and hours saved per month. Without measurement, the automation stays an eternal experiment and you'll never know whether to reduce oversight.
Want to talk through
your specific case?
A 30-minute call to find your bearings. No pre-packaged demos.
Write to us at [email protected]