Choosing what to automate
What should we automate with AI?
Last updated
Justas Butkus is a fractional AI officer based in Vilnius, Lithuania, working with mid-market companies and scale-ups across the UK, EU and US. He builds and operates the production AI systems he advises on, and takes no commission, referral fee or revenue share from any vendor.
Short answer
Automate the process that is frequent, expensive and disliked. Frequency matters more than difficulty: a two-minute task done four hundred times a month is worth more than an hour-long task done twice a year. If a process is broken, automation makes it faster and no better.
Which processes are usually worth it?
| Process | Usually worth it? | Why |
|---|---|---|
| Sorting and routing inbound email or forms | Yes | High volume, rules already exist in someone's head |
| Extracting data from documents into a system | Yes | Pure re-keying, high volume, easily measured |
| Answering repeated internal questions | Yes | The answer already exists, finding it is the cost |
| First-draft replies for a person to edit | Often | Works when a human reviews; risky when they do not |
| Chasing and reminders | Often | Gets skipped when busy, which is when it matters most |
| Anything requiring judgement never written down | No | You cannot automate a rule nobody has articulated |
| Low-volume, high-stakes decisions | No | Not enough repetition to pay back, too much downside |
| A process people already work around | No | Fix the process first, or you automate the workaround |
Why does frequency matter more than difficulty?
Because the return is the product of the two numbers, and frequency is usually the larger one.
A task taking two minutes, done four hundred times a month, is roughly a hundred and sixty hours a year. A task taking a full day, done twice a year, is sixteen. The first is almost always the better candidate even though it feels less significant, and the second is the one people instinctively pick because it feels painful.
This is also why the interesting process is usually the wrong one. Interesting tends to mean rare and complicated, which is the worst combination for automation.
What disqualifies a process?
- Nobody can write down the rule. If the decision lives in one experienced person's judgement and has never been articulated, capturing it is the work, and it is a different project.
- The data is inconsistent. If the same field means different things in different systems, any system will learn that inconsistency faithfully.
- It is broken already. Automating a broken process produces a faster broken process, and now it is harder to see.
- Nobody will accept responsibility for it running. Without a named person, it will stay in testing indefinitely.
How do you know it worked?
Decide this before you build, in a number, or you will end up arguing about impressions.
- Time: the same volume handled in measurably fewer hours.
- Volume: more handled with the same people, if that is what you actually need.
- Latency: how long something waits before a human sees it, which is often the real complaint.
- Error rate: measured against what people were achieving before, not against perfection.
That last point matters. The comparison is the current human process with its real error rate, not an imagined flawless one.
Frequently asked questions
What should a company automate with AI first?
The process that is frequent, expensive and disliked. Usually that means sorting inbound work, extracting data from documents, or answering the same questions repeatedly. Frequency matters more than how difficult the task is.
What should we not automate?
Anything requiring judgement nobody has written down, low-volume high-stakes decisions, and any process people already work around. That last one is a process problem, and automation makes it faster without making it better.
How do we measure whether the automation worked?
Decide before building, as a number: hours for the same volume, volume with the same people, how long work waits before a human sees it, or error rate compared with what people actually achieved before.
Is a two-minute task worth automating?
Often yes, if it happens constantly. Two minutes done four hundred times a month is about a hundred and sixty hours a year. The long, rare task feels more painful but usually returns less.
If you have a shortlist
Ranking it takes an afternoon and the method is on this page. A second opinion is a short conversation.