How to turn a repetitive task into an AI workflow: a 6-step method, with a worked example

First, pick the right task

Good first candidates share four traits:

• It repeats, weekly or more often.
• It takes real time, enough that you'd notice getting it back.
• It follows a pattern: similar inputs and a similar output each time.
• A mistake is cheap to catch before it reaches a customer.

Weekly reports, lead clean-up, first-draft replies to common questions, meeting follow-ups and client update emails usually qualify. Anything that moves money or makes final decisions about customers doesn't, at least not as a first project.

Throughout this guide we'll use one example: the Monday report. An operations manager at a 15-person software company spends about three hours every Monday pulling numbers from the CRM and two spreadsheets into a report that, in her words, nobody reads past the first page. Illustrative example

Step 1: Define the outcome

Write one sentence that describes what changes, in terms the reader of the output would recognise. Not "automate the report", but:

Five bullet points in the team chat every Monday at 9:00, with the numbers used shown underneath and a flag when something looks off.

That sentence already rules out most of the work people wrongly automate: charts nobody reads, a 12-page deck, a dashboard to maintain.

Step 2: Map how it's done today

List every step in order, who does it, and in which tool. Be honest about the steps nobody likes to admit:

1. Export last week's deals from the CRM.
2. Copy the leads tab from the marketing spreadsheet.
3. Copy the revenue tab from the finance spreadsheet.
4. Fix date formats and remove duplicates by hand.
5. Add up totals by channel.
6. Compare with last week and look for anything odd.
7. Write the summary and send it.

Mark where it breaks or slows down. Here, steps 4 and 6 take most of the time, and step 5 is where errors creep in.

Step 3: Write the spec

A spec is half a page that any person or tool could follow. It has five parts:

OUTCOME   5 bullets every Monday 9:00 + numbers used + flags.   (illustrative)
INPUTS    deals.csv (CRM export), leads and revenue tabs.
STEPS     clean → total by channel → compare with last week → write bullets.
RULES     Calculate every number; never estimate. Show the table of
          figures used. If data is missing, say so instead of guessing.
          Don't send anything; put the draft in the shared folder.
DONE      5 bullets, a figures table, and any flags, ready for a person
          to approve in under 5 minutes.

The rules line does most of the work. In our own tests, "calculate and show the figures" was the difference between numbers you can trust and numbers you can't. See that test.

Step 4: Decide where AI helps, and where a simple rule is better

In the Monday report, cleaning and totals are rules; spotting "LinkedIn brought 8 leads but no meetings" and writing the bullets is where AI earns its place.

Step 5: Add checks and a person

• Show your working: the figures table goes under the bullets, so a person can check in seconds.
• Mark gaps: missing data is flagged, never filled in.
• Approve before it goes out: the draft lands in a folder or channel; a person sends it.
• Log every run and get alerted when one fails. One team reports a $1,400 bill in one hour after an agent looped on a repetitive task with nobody watching.

Step 6: Choose the tool, and build one step at a time

Now the tool question is easy, because you know what it has to do. For a weekly report from files you already have, the AI plan your team pays for (Claude, ChatGPT or Gemini) is usually enough; a connected automation tool helps if the data must be pulled from several systems on a schedule. Why we pick the tool last.

Then build in small steps, not all at once:

1. Get the clean-up and totals right on last week's files.
2. Add the bullets and flags.
3. Run it alongside the manual report for one or two weeks and compare.
4. Only then make it the real report, with the approval step kept.

What is it worth? A worked calculation

An honest return calculation uses your own numbers. Here it is for the example, with every input stated so you can swap in yours. Illustrative

• Hours back: (3 − 0.5) × 48 = 120 hours a year.
• Value of that time: 120 × $50 = $6,000 a year.
• Payback on a $750 set-up: $750 ÷ (2.5 hours × $50 = $125 a week) = 6 weeks.

Two cautions: hours saved only become value if they're spent on something better, and the first few weeks include fixing and tuning. A real project should also count any tool subscription you don't already have.

Mistakes to avoid

• Handing over the whole job at once. Users describe this as where sessions go off the rails; plan first, then one step at a time.
• Automating a broken process. If nobody reads the report, fix the report before you automate it.
• No checks, no owner. Every workflow needs someone who looks at the output and owns it when it breaks.
• Starting with the tool. Picking a platform first tends to shape the work around the tool instead of the outcome.

Common questions