How to Work Out Which Tasks Are Actually Worth Automating

The usual advice about AI automation is to "find repetitive tasks". That is true and almost useless, because most of a working week looks repetitive from the outside. The question is which repetition is worth removing.

Write the process down first

This is the step everyone skips, and skipping it is why most automation projects quietly fail. If a process only exists in someone's head, automating it means encoding a version of it that nobody has ever checked.

Take one process. Write it as numbered steps, and mark each step as one of three things:

• Manual — someone does it by hand, the same way each time.
• Rules-based — the decision follows a rule you could write down.
• Judgement — it depends on context, relationships, or risk.

The first two are candidates. The third is not, however tempting.

Score it honestly

For each candidate step, you want three numbers: how often it happens, how long it takes, and what it costs if it goes wrong.

Frequency times duration tells you the prize. The cost of being wrong tells you how much checking you need to build around it. A task that happens forty times a week and takes two minutes is worth more than one that takes an hour but happens monthly — and if getting it wrong just means someone fixes a typo, you can move fast.

The tasks worth doing first sit in the same place every time: high frequency, rules-based, low consequence.

Keep the human where the judgement is

The most common mistake is not automating too little. It is automating the wrong half.

A support process might have five steps: read the message, work out what the person actually wants, find the answer, write the reply, send it. AI is genuinely good at the fourth. It is unreliable at the second, which is where the relationship lives. Automate the drafting and leave the understanding with a person, and you get most of the time back with none of the risk.

Decide what "working" looks like before you start

Pick the measure before you turn anything on. Time saved per week. Response time. Error rate. Anything, as long as it is one number you could actually check in a month.

Without that, you will end up defending the tool rather than evaluating it — and the honest answer to "did this help?" is often "we do not know", which is an expensive place to be.

What to leave alone

Some work looks automatable and is not:

• Anything where the relationship is the product.
• Anything where being wrong is expensive and hard to notice.
• Anything you do rarely enough that you will forget how the automation works.

Do it on one process

Pick the single process that costs you the most time relative to its value. Map it. Automate one step. Measure it for a month. Then do the next one.

That is slower than it sounds like it should be, and it is the only approach that survives contact with a real business.