Process19 Sep 2026

Hand AI the boring task first

By Alan Goh · 2 min read

The test

A good first task for AI

  • It comes back every week.
  • You could write its steps down.
  • You can check the answer in two minutes.
Short answer

The best first task for AI is a routine one: it comes back every week, you could write its steps down, and you can check the answer. The impressive one can come later.

When people choose their first task for AI, it’s tempting to pick the most impressive one: the pricing decision, last week’s difficult customer, the job that would prove the whole idea.

Those tasks are hard for anyone, and a tough first experience can leave people thinking AI isn’t ready for their business. Usually, it was the choice of task.

A good first task

Three things, and all three matter.

It comes back every week. A job that returns on schedule pays you back on schedule, and every run teaches you something about how the tool behaves.

You could write its steps down. If you can explain it to a new hire in ten minutes, you can explain it to AI. Tasks that live as instinct in one person’s head are better saved for later.

You can tell whether the answer is right. Quick checking is what turns AI’s work into time saved.

The weekly summary report. The quotation built from the same price list every time. The reply to the question your staff answer five times a day. Routine, every one of them, and that is what makes them good places to start.

Tasks to save for later

A crisis. High stakes are the hardest place to learn how a new tool behaves.

Something outside your own expertise. You’ll want to be able to check the answer.

The most impressive thing you can think of. Impressive tasks tend to be rare, vague and hard to check. They make a great second or third project.

Support first, run later

Ask one more question of any task: will AI support this, or run it?

Supporting means AI produces the draft and a person checks it before anything leaves the building. Running means it acts on its own, every time.

Start with support. You learn how the tool behaves while a person still checks its work. When it has been reliably right for a while, you can decide whether it should run, and by then you’ll know what to watch for.

Start with information you’d be comfortable sharing

For the first task, choose work that doesn’t involve IC numbers, bank details or other sensitive customer data. Learning the tool comes first; handling sensitive data well is a separate lesson, best taken once you’re comfortable.

The test

Ask three things of the task: Does it come back every week? Could I write its steps down? Could I check the answer in two minutes?

Three yeses make a great first task.

In your department

  • Accounts & Finance: The weekly ageing report, or the payment reminder built from the same template, makes a good first task: routine, written down, and quick to check against the ledger.
  • Sales & Marketing: The quotation built from the same price list every time is a strong place to start. Check it against the list before it goes out.
  • HR & Admin: Start with the reply to the leave or claims question staff ask five times a day. Leave anything with IC numbers or salaries for later.

Want to try this with your team?

Tell me what your team is working on, and I'll suggest where to start. I read and reply to every message myself.

More notes

  • 09 Sep 2026

    Five skills for getting good results from AI

    Getting good results from AI comes down to five skills: break the task down, give it the context, set the limits, refine the draft, and judge the result. When an answer disappoints, the missing skill tells you what to fix.

    Read it →
  • 02 Sep 2026

    Keep your AI instructions in a file

    The instructions you give AI for a repeated task are worth keeping in a plain document. Use the file every time, update it whenever you correct a result, and the task gets better with every run, whoever on the team runs it.

    Read it →
  • 19 Aug 2026

    Split AI work into asking, doing and checking

    AI work goes better with three roles kept apart: you decide and write down what good looks like, a fresh AI session does the work from that instruction alone, and someone with no stake in the result checks it against your description.

    Read it →

All the notes →