Automate the task that takes up the most time and follows the same steps every time you do it. To find it, track a full week of work instead of sitting down to brainstorm ideas. Most businesses I talk to already have one or two of these tasks, and they're usually more boring than whatever the owner was planning to automate.
This is the same process I use with clients before I build anything for them. It's part of a bigger guide on AI automation for small business, if you want the full picture first.
Why brainstorming picks the wrong tasks
When people sit down and ask themselves what they should automate, their gut gives them two kinds of bad answers.
The first is a task where building the automation takes longer than doing the work ever did. There's an xkcd comic with a whole chart of how much time you can spend automating something before it stops paying off, and most people never check. Developers are the worst at this, me included, because building things is fun. Spending ten hours on an automation that saves one hour sounds great when you enjoy the ten hours.
The second is a task that isn't worth anything. It's something you do twice a year, or something that wouldn't change much if it disappeared. Automating it feels productive, but you get almost nothing back.
Both mistakes come from guessing. So stop guessing and collect some data.
Track a week of work
For one week, write down everything you do, in as much detail as you can stand. Use time blocks: what you worked on and how long it took. If you have a team, ask them to do the same.
A week is long enough to see where your time actually goes, and it's almost never where you think it goes. You might think email takes you half an hour a day and find out it's closer to two hours.
When I run this with a client, I give their team a way to log their work so it doesn't become a chore. Depending on the business, that's a spreadsheet or a small app where each person records what they're doing, and sometimes an automatic time tracker. The tool matters less than the detail. "Admin, 2 hours" tells you nothing, but "copied new orders from email into the spreadsheet, 40 minutes" tells you exactly what to automate.
What a good first task looks like
Once you have a week of data, look for the tasks that eat the most time and come back over and over with the same steps. The more manual and boring a task is, the easier it is to automate.
Here's how I'd judge each task on the list:
| Question | Automate it | Leave it for now |
|---|---|---|
| How much time does it take each week? | Hours | A few minutes |
| How often does it happen? | Daily or weekly | A few times a year |
| Are the steps the same every time? | Yes, you could write them down | They change with every case |
| What happens if it goes wrong? | It's easy to spot and fix | A customer or the business gets hurt |
The tasks that land in the left column on all four questions are your best candidates. Pick the one that takes the most time and start there.
Tasks I tell clients to leave alone
Some tasks look like good candidates and aren't.
If something isn't very repetitive, I'll usually say leave it. A task that comes up once a month can take longer to automate than to do by hand for years, and it changes a little every time it comes up.
I'd also leave anything that needs complex reasoning or real creativity. AI can help with parts of that work, but people still do it better, and an automation built around it will need someone checking its output anyway.
Sensitive tasks are the last group. If a mistake would hurt a customer or cost the business money, a person should be doing it, or at the very least approving what the automation produces before it goes anywhere. I cover how that approval step works in how to make AI automations reliable.
Write the steps down before you build
Before you automate anything, write the process down step by step, as a list or a simple flowchart. Then do the task by hand once or twice while following your own instructions.
If you can't do it manually, you won't be able to automate it. And if you can't explain every step in detail, you don't understand it well enough yet. People who have done a task for years run on muscle memory and skip over small steps without noticing, and those small steps are exactly what breaks an automation later.
I still haven't automated the thumbnails for my YouTube videos, and this is why. Every thumbnail is different and I don't have a repeatable process for making them. Until I do, there's nothing to automate.
My video editing was the opposite. I'd edited dozens of videos by hand before I wrote a single script, so I knew every step. I automated the audio cleanup first, then added silence trimming, then captions, one small piece at a time. Each step was easy to automate because I already knew exactly what it involved. My short videos on writing the process down and getting granular go into this in more detail.
You don't need to automate all of it
A lot of people treat automation as all or nothing, and if they can't automate a task completely they don't bother with it at all.
If someone spends 10 hours a week on a task and you automate 80% of it, they get 8 hours back every week. The remaining 20% might be the part that needs a human anyway, like checking the result or handling the odd case that doesn't fit.
Most tasks have a big repetitive chunk that plain code can handle without any AI, plus a smaller part that needs judgment. Automate the chunk first. I talk about this in automation isn't all or nothing, and the split between code and AI is the whole subject of AI agents vs workflows.
How Mastersealer picked its first task
Mastersealer is a landscaping business in Vaughan, Canada, and one of my clients. Johnny, who runs it, asked me to automate their quoting.
It was a good first task. Johnny was writing every quotation by hand and emailing each one to the customer as a PDF, and he was doing that constantly. The steps were the same every time: take the customer's details, fill in the quote, make the PDF, send it. Nothing in that chain needed judgment, so I built it as a small app in plain code with no AI in it at all.
It took me somewhere between 10 and 20 hours to build. Johnny now fills in a short form and the app does the rest, though he still checks each quote before it goes out. Together with automating their invoicing, it saves them more than 10 hours a week. The full case study has the rest of the story, including the $430k CAD in quotes that had been landing in spam.
Then do the next one
Once the first automation is running and you trust it, go back to your week of tracking and pick the next task. Automate one part of the business at a time. Trying to do everything at once is how you end up with five half-finished automations and nothing that works.
If you're doing this yourself, block out dedicated time for it, an evening or a weekend, so it doesn't get squeezed between everything else. I explain the whole approach in what to automate and how to actually automate your business with AI.
If you'd rather someone else ran the tracking week and built the automations, that's what I do. Here's how I work, and if you want to know what it'll cost to keep running afterwards, see what AI automation costs.
