Which parts of your business should AI handle?

The mistake I watch people make is asking "what could AI do?" instead of "what should it do?" Automating the wrong task costs more than doing it by hand, because now you're babysitting a bad robot on top of the work. The fix is to inventory the work first, then score it, so you start where the payoff is real.

If you came from the video, that's the idea. Here's the part that doesn't fit in sixty seconds: the exact three-number score I use, one business scored line by line, and the traps that turn a "good" candidate into a mess.

List the recurring work first

Open a blank doc and write down the tasks your business does over and over. Answering the same five questions. Drafting captions. Sorting inbound leads. Summarising calls. Chasing invoices. The test is simple: if you've done it more than a handful of times and you'll do it again, it goes on the list. Leave off the one-offs, because repetition is what pays back the setup cost. Aim for eight to fifteen tasks, and if you're unsure whether something is one task or three, split it.

Score each task on three numbers

For every task, rate three things from 1 to 5, then multiply them. The multiply matters: a task has to score decently on all three to rise, so one big number can't drag a bad fit to the top.

Multiply the three, sort high to low, and you have a queue, not a vague sense you "should use AI more."

Watch it work on one real business

Abstract scoring is easy to nod along to and hard to trust, so let's run it on someone specific. You own a small Pilates studio: two instructors, forty-odd regulars, and you do everything that isn't teaching. Here's the list, scored honestly:

• Answering "class Tuesday at 6?" style DMs: 5 x 4 x 5 = 100. Same handful of questions, all day, all answerable from your schedule.
• Writing the weekly class-update email: 4 x 4 x 4 = 64. Repetitive and draining, and a draft from your notes gets you most of the way.
• Chasing lapsed members who haven't booked in a month: 3 x 5 x 3 = 45. You hate it so you skip it, but AI drafts and you send.
• Deciding whether to add a Saturday class: 1 x 2 x 1 = 2. A judgment call about your money and your weekends. It stays with you.
• Teaching the class: doesn't go on the list. It's the thing customers pay for.

The DM answering wins by a mile, and for the right reason: it's constant, it wears you down, and it's exactly the look-it-up-and-reply work AI is good at now. Start with the weekly email because it "feels more like marketing" and you burn your setup energy on the second-best task while the top one keeps costing you.

Copy the scorecard and run your own

Here's the sheet I use. Drop it into a doc or spreadsheet, fill in your tasks, and let the Score column sort your queue.

Task | Frequency (1-5) | Pain (1-5) | AI fit (1-5) | Score
--------------------------------------------------------
[task 1] |   |   |   | =F x P x Fit
[task 2] |   |   |   |
[task 3] |   |   |   |
[task 4] |   |   |   |

Rating guide:
- Frequency: daily=5, weekly=3, monthly=1
- Pain: dread it=5, mildly annoying=3, fine=1
- AI fit: text + repetitive + low-stakes=5,
          judgment + relationship + high-stakes=1

Sort by Score, highest first. Keep a human review step on
anything that touches money, customers, or your reputation.

Not sure how to rate AI fit? Ask the model to do the task once, small, with fake data. If the output only needs a glance before sending, that's a high fit. If you'd rewrite half of it, it isn't.

Automate the top one, not all of them

This is where discipline beats enthusiasm. Take your single highest-scoring task and set up AI for just that one. For the studio, that's a booking-question responder wired to the class schedule. The build your first AI agent guide walks through the four pieces it needs, and my MCP connectors guide covers hooking it to your real calendar and inbox so it isn't guessing. Run it, watch it, fix what it gets wrong, and only then move to number two. Slow and proven beats a dozen half-built automations you don't trust and quietly stop using.

What to watch for

The score points you at a candidate. It doesn't guarantee the automation works, and here's where good on paper turns into a mess.

People inflate AI fit because they want the task to qualify. The tell is the review step: if checking the output takes almost as long as doing the task yourself, the fit was lower than you scored it. The honest test isn't "can AI produce something," it's "can AI produce something I only have to glance at."

High frequency plus low stakes is a trap when "low stakes" isn't true. Answering booking DMs looks harmless until the agent tells someone there's a 6pm class that got cancelled, and you've lost a customer over a one-line reply. Anything customer-facing needs a memory of the current truth, which is why the studio's responder reads the live schedule, not a week-old snapshot.

The task you resent is not always the one to automate first. Pain is on the scorecard, but a once-a-month task you dread loses to a daily one you tolerate, because the daily one steals more of your life. Trust the multiply over your gut.

And the quiet one: some of your list shouldn't be automated, it should be deleted. Before you hand a task to AI, ask whether it needs doing at all. A weekly report nobody reads doesn't get an automation, it gets cancelled. AI is good at helping you do useless work faster, and it's still useless.