What should you automate first?

Why “start with what annoys you” is the wrong instruction

It is the most common advice, and it is backwards. Annoyance measures how a task feels, not how much of your week it takes. The jobs that irritate owners most tend to be short, infrequent and stuffed with judgment: the awkward client email, the quarterly filing, the one supplier who never replies. Those are low hours and high difficulty, which is the worst possible combination for a first automation.

The tasks that actually consume the week rarely register as annoying, because they have been absorbed. You stopped noticing them years ago. Rewriting the same three paragraphs for a new proposal, pulling the same numbers into the same weekly summary, retyping a booking into a second system, chasing the same two documents from every new client. None of it feels like a problem. All of it is the problem.

The work worth automating first is invisible to you, precisely because you have done it so many times.

The three numbers that decide the order

Score every candidate task against three things. You do not need a spreadsheet to start, but you do need to write the numbers down, because your memory of them will be wrong.

• Hours per week. Not per instance. A four minute task done thirty times a week outranks a two hour task done monthly, and almost nobody ranks them that way by instinct.
• Variation. How different is each run from the last one? If you could hand somebody a one page instruction and get an acceptable result, variation is low. If the instruction would need to say “it depends,” variation is high.
• Judgment cost. If this comes out wrong, what does it cost to fix? A mistyped social caption costs a minute. A mispriced quote costs a client.

Your first automation is the task that sits highest on hours, lowest on variation, and lowest on judgment cost. Rank the list on all three and the order stops being a matter of opinion. That combination is also what makes the difference visible fast, which matters more than people expect: the first one has to work, or the second one never gets built.

How do you count the hours honestly?

Count the whole cycle, not the visible step. Writing the weekly client update is not the fifteen minutes of typing. It is opening four tabs to gather the numbers, the typing, the wait while somebody confirms a figure, the second pass after they reply, and the twenty minutes of scattered attention it costs to pick your real work back up afterwards.

Track it live for one week rather than estimating from memory. A note on your phone with a start and stop time is enough. Most owners undercount recurring admin by a factor of two or three, and the gap is almost always the coordination around the task rather than the task itself.

What does a good first automation actually look like?

It has a clear trigger, a defined input, and one output you can look at and immediately judge as right or wrong. Nothing about it should require you to be present at a particular moment.

Concretely: a new enquiry arrives, and within two minutes the person gets a reply with the right next step, the details are already in your CRM, the intake documents have been requested, and you have a one paragraph summary waiting whenever you next look. You did not touch it. You can still read every part of what happened.

That is the shape to aim for. Not a system that runs your business, a system that finishes one repeating job end to end while you are somewhere else. If you want the fuller version of this, the free 10 AI Agents guide names ten of these jobs by function and says what each one replaces.

What should you not automate first?

Three categories, and they are worth stating plainly because skipping them is how automation projects earn a bad reputation inside a business.

• Expensive to reverse. Anything touching pricing, contracts, payroll or money leaving the business. These can be automated later, behind a human checkpoint, once the pattern is proven on lower stakes work.
• Different every time. If two runs of the task share almost no structure, you are not automating a process, you are automating a decision. That is a later phase and a different design.
• Never documented. If you have never written down how you do it, an agent cannot do it either. This is the most common blocker, and it is also the cheapest to fix: describing the task properly is usually an hour of work that pays for itself even if you never automate it.

How do you know it worked?

Measure the same three things before and after, at the same time of week:

1. Hours on that task per week. The headline number, and the one you already counted.
2. Elapsed time from trigger to finished output. Often the bigger win. A task that took you twenty minutes but sat in a queue for two days now finishes in four minutes, and the client notices that long before they notice anything else.
3. Interventions. How many times did a human have to step in and correct it? If hours went down but interventions went up, the work did not disappear. It moved, and it will drift back.

Then guard the recovered time deliberately. Reclaimed hours refill on their own if nothing is put in their place, and an owner who wins back six hours a week and spends all six on new admin has proven the mechanism works and gained nothing from it.