The short answer: Automate the task you repeat most often that requires the least judgment. For most small businesses that is one of three things: answering the same inbound question, sending the same follow-up, or moving the same information from one place to another by hand. Pick one. Document how you do it today. Then, and only then, look at tools.

The order matters. Most automation projects fail because they start with software and work backward to a problem.

Why businesses pick wrong

The usual story goes like this. The owner reads about an AI tool, signs up for a trial, and spends a weekend connecting it to things. By Monday there is a workflow that sends a Slack message when a form is submitted. It is impressive. It also solves nothing, because nobody was losing sleep over form notifications.

Meanwhile, the owner is still personally answering "do you work with businesses my size?" eleven times a week, still typing the same onboarding email with the client's name swapped in, and still copying invoice totals from one app into a spreadsheet every Friday.

The tool got picked before the problem did. That is backward. In the Navy, you diagnose before you treat. Treating symptoms you have not examined wastes supplies and sometimes makes things worse. Automation works the same way.

Three questions before you automate anything

Here is the diagnosis I run with clients before we talk about a single tool.

1. What do you do more than five times a week that feels the same every time?

Write them down. Do not filter. Answering the same email. Scheduling. Chasing a signature. Formatting a report. Posting the same kind of update. Reminding someone of something. You will find more than you expect.

2. For each one, how much judgment does it take?

Rate it honestly. Replying to "what are your hours" takes none. Deciding whether a lead is a good fit takes a lot. Drafting a proposal takes some judgment at the start and the end, and almost none in the middle.

3. What breaks if it is done wrong?

A late reminder is annoying. A wrong invoice costs money. A tone-deaf reply to an upset customer costs the relationship. The lower the stakes, the better the candidate for early automation.

The first thing to automate is at the intersection: high frequency, low judgment, low stakes. That is where you get the time back fastest with the least risk.

The one I see most

Across most of the businesses I work with, the winner is the same: inbound questions that already have an answer.

Every business gets the same fifteen questions. Pricing, availability, whether you serve a certain area or customer type, what happens after someone signs up, how to reschedule. The owner answers each one fresh, from memory, slightly differently every time. That inconsistency is its own problem, and it is a brand voice problem as much as an efficiency problem.

Write the fifteen answers once, in your voice. Put them somewhere your team and your tools can read them. Now a form, an email assistant, or a chat tool can give the same correct answer every time, and the owner reviews the ones that fall outside the fifteen. You have not replaced anyone. You have stopped paying a skilled person to retype the same paragraph.

The second most common winner is follow-up. Someone fills out a form, books a call, or buys something, and the next three touchpoints are identical every time. That is a sequence, and a sequence should run itself.

Automate repetition. Support judgment. Preserve accountability.

That line is the whole philosophy. Automation is for the repeated middle of a process. The decisions at the beginning and the review at the end stay with a person.

A good test: can you write down the exact steps, and would two different people following those steps get the same result? If yes, it is a candidate. If the answer is "well, it depends," you have found a judgment call, and judgment calls get supported with better information, not replaced with a rule.

I keep a human approval checkpoint in every automation I build, even the simple ones. Not because the tools are unreliable, but because a business should always be able to answer "who decided that?" with a name.

A short story about doing it the hard way

When I migrated 674 files out of three and a half years of ChatGPT history, my first instinct was to find a tool that would sort them for me. I spent an evening evaluating options. None of them knew what my categories should be, because I had not decided yet.

So I did the first fifty by hand. By file thirty I had a pattern: identity, customers, offers, processes, stories, standards. By file fifty I could write the pattern down. Then I handed the remaining 624 files and the written pattern to an AI tool and reviewed its sorting in batches. It took a fraction of the time, and it was right most of the time, because I had done the judgment work first and given it something to follow.

That is the sequence. Do it by hand until you can write it down. Then automate what you wrote down. Keep reviewing.

What not to automate first

A few things that look like good candidates and are not:

  • Anything customer-facing that touches an upset person. The cost of getting it wrong is too high to be your first project.
  • Anything you have not done manually at least a dozen times. You do not understand it well enough yet.
  • Anything you are automating because a tool made it easy. Easy to build is not the same as worth building.
  • Your content. AI can draft, but if you automate publishing without review, you will eventually publish something that is not you. Content is a Layer 3 problem; it comes after you have captured your voice.

An example from a service business

A home services company I worked with wanted to "automate their marketing." That is not a task; it is a wish. When we ran the three questions, the real list looked like this:

  • Replying to quote requests with the same four qualifying questions (12 times a week, zero judgment, low stakes)
  • Sending appointment reminders (30 a week, zero judgment, medium stakes if missed)
  • Writing up the job summary after a visit (8 a week, some judgment, medium stakes)
  • Deciding which jobs to bid on (5 a week, high judgment, high stakes)

We automated the first two in an afternoon with tools they already paid for. The third got a template and an AI draft with the tech reviewing it. The fourth stayed with the owner, who now had five more hours a week to think about it.

Where to start

Take the AI Readiness Assessment if you want a structured read on whether your business has the information, processes, and safeguards to automate well. Or skip straight to the three questions above with a notepad. Either way, you are looking for the task that is boring, frequent, and low risk. That is your first automation, and it is probably not the one the software ad told you about.