How do you pick the first process to automate with AI in a small business?

Why "where do we even start" is the real first problem

Most small businesses don't struggle for lack of ideas — ask around the office and you'll get a dozen. The struggle is that most AI projects that stall or never ship started with the wrong one: a process too rare to matter, too messy to trust, or too risky to hand over. Picking well the first time is what makes the second and third process possible.

The four-question scorecard

• Does it repeat at least weekly? A task done twice a year is never worth building around, no matter how painful it is on the day. A task done every week, or every day, pays back the setup fast.
• Is the data already in one system, not scattered across inboxes and paper? A process whose numbers live in one place is quick to automate well. A process that depends on someone's memory, a sticky note, or three different inboxes is quick to automate badly.
• Is a mistake cheap and easy to catch? Look for work where a wrong first draft is easy for a person to review and fix, not a one-way action like sending money or a message that can't be pulled back. Start there; add a human check step everywhere the stakes are higher.
• Does someone already do this by hand, every week, right now? If a real person is already doing the task, there's a real before-and-after to measure. If the "process" doesn't really exist yet, there's nothing to compare the result to.

What usually wins the scorecard

A recurring report that gets rebuilt from the same numbers every week. A status update that gets re-typed from one system into another. Matching numbers between two systems that don't talk to each other. Answering the same handful of questions over and over, whether from customers or from inside the company.

What usually loses, even when it feels urgent

Anything that moves money or sends something outward with no human check. Anything that happens only once or twice a year, no matter how painful. Anything where the "process" really lives in one person's head and has never been written down — that needs to become a real process before AI can help it, not the other way around.

What happens after you pick one

Prove it on that single process first — a small, visible win the whole team can see — before taking on a second. Each month after that, review what's working, fix what isn't, and pick the next process using the same four questions.

A short worked example

Say a company is weighing two ideas: a weekly report a manager rebuilds by hand from the same two systems every Friday, and a once-a-year contract-renewal push that everyone dreads. The report scores well on all four questions — weekly, one place, low-stakes to review, and someone already does it today. The renewal push fails the first question outright: once a year isn't often enough to be worth building around, no matter how painful that one week feels. The report wins, even though the renewal push is the one people complain about most.

Why the obvious answer isn't always the busiest-looking task

It's tempting to start with whatever feels most overwhelming right now. But a process that's merely unpleasant once a year will never pay back the time spent automating it, while a quieter task that repeats every single week adds up fast — and gives you an early, visible result to point to before asking the team to trust a second one.

What you own

The scorecard itself, and the process documentation that comes out of using it, are yours either way — useful even if you end up building the first tool yourselves.