Skip to main content

Operations

AI automation for a small business: what to automate first

AI automation for a small business, decided on paper. A three-question score to pick the first task, six worked examples, and what to leave alone.

A woman in a denim shirt at a whiteboard in a breeze-block storeroom, drawing out a process grid, a work lamp throwing hard green-white light across the board.

Forty things eat your week. The only question that matters is which one to hand over first.

Search for AI automation for a small business and you'll get lists of tools instead. Ten of them, with the writer's own product at number one. A list of tools answers a question you haven't got yet.

Here's a way to answer the one you have. On paper, fifteen minutes, no software and nobody's sales call.

The score

Take one task. Ask three questions, score each one, add them up.

Score one task out of nine
Ask about the taskScore 3Score 2Score 1
How often does it happen?Every dayEvery weekMonthly or less
What does it cost when it's wrong?An apologyA callback or a creditA lost customer
Are the steps written down?Written so another person could follow themRoughly known by two peopleLiving in one person's head
The cost row runs backward on purpose. You want to make your first mistakes where being wrong costs you an apology, not a customer.

Add the three. Nine is where you start. Three means leave it.

Most things come out at six or seven, which is the useful middle. A seven with the steps already written down is worth starting on this month. A seven where the one is the written-down question isn't a task at all. It's a writing job wearing a task's clothes.

The gate that beats the score

One rule overrides all three. If the steps aren't written down, the score doesn't apply yet.

A tool can only follow a process you can describe. Hand it a process that lives in your estimator's head and you don't get automation. You get a confident guess at what the process might be. So write the steps first, which starts with getting them out of the head they live in. The prompt and the five-step check for doing that take an afternoon.

Figure 1 / Decision

If the steps are not written down the score does not apply yet, whatever the three questions add up to.

The gate in front of the score

  1. Needs a person

    Are the steps written down?

    • No

      Affirmed

      The score does not apply yet. Write the steps first.

    • Yes

      Outcome

      Score it

      Add the three. Nine is where you start. Three means leave it.

  • Needs a person
  • Affirmed
RunSimpler

Six worked examples

These are illustrations of the method, not accounts of work we've done. Run your own numbers.

The follow-up message after a completed job. Daily is 3, an apology is 3, and the steps are obvious once somebody types them out, so 3. Nine, and it's where a service business should start. The input is the job note. The output is four sentences.

Maintenance agreement renewals. Seasonal, so the frequency score is low. But the steps are simple, and a clumsy one costs you an apology. Good second task, and worth timing for the shoulder season instead of July. The seasonal version of this whole question is in AI for HVAC companies.

The recall list at a dental practice. Weekly, and the cost of a wrong message is higher than it looks. Let a tool draft the message and keep a person sending it, with no patient detail going into any tool. The reasons are in what your team can safely put into ChatGPT.

The monthly client report at an agency. Monthly, expensive when wrong, and the steps are probably written down somewhere, because a client agreement describes them. Draft yes, send no. More on where this helps and where it hurts in AI inside a small agency.

Turning a shop job into a written changeover procedure. Not automation at all. It's what makes the next three possible. That's why it's on the list.

Quote chasing. Daily is 3 and an apology is 3, but write the steps out honestly and the third score is usually a one, because "look at the list, send the message" leaves out how the list gets made. The hard part was never the writing. It's that nobody looks at the list on a busy week. That's the bit you're automating.

Leave these alone for now

Not forever. For now, while you're learning where the tool is reliable.

  • Anything that reaches a customer with a price or a promise in it and nobody reading it first.
  • Anything that diagnoses equipment. A model number lookup is fine. A diagnosis belongs to a licensed person, and that line doesn't move.
  • Anything in the hiring or firing path.
  • Anything where you couldn't tell a wrong answer from a right one by looking. If you can't check it, you can't own it.
  • Anything that lands on a customer's phone at 11pm because a rule fired. Put a window on it.

Four of those five come down to the same guard, and it's a person sitting between the tool and the customer.

The fifth is a clock. Set a sending window and nothing fires outside it. If the tool has no window setting, the tool has just told you what it is.

Figure 2 / Boundary

Nothing comes off the leave-alone list until three weeks of a person pressing send have given you nothing to fix.

What to hand over and what to leave

Hand over now

  • The follow-up message after a completed job

    Daily, cheap when wrong, steps obvious. Where a service business should start.

  • Quote chasing

    The steps are "look at the list, send the message".

  • Maintenance agreement renewals

    The good second task, timed for the shoulder season.

Leave these alone

  • Anything that reaches a customer with a price or a promise and nobody reading it first

  • Anything that diagnoses equipment

    A model number lookup is fine. A diagnosis belongs to a licensed person.

  • Anything in the hiring or firing path

  • Anything where you couldn't tell a wrong answer from a right one by looking

    If you can't check it, you can't own it.

  • Anything that lands on a customer's phone at 11pm because a rule fired

    Put a window on it.

The lineFree tool, a person presses send, three weeks, then decide.

RunSimpler

Start free, and send it yourself

The order that works is boring. Free tool, person presses send, three weeks, then decide.

Buy the platform first and you're paying monthly for a workflow you haven't proved, and configuring it before you know what you want it to do.

Automating a twelve-step process gives you a faster twelve-step process. Look at the steps before you make them permanent. Ask out loud why each one exists. Two of the answers will surprise you.

What "automated" looks like at eleven people

Three stages, and you can't swap them around.

  1. DraftedThe tool writes it, a person reads it, the person sends it. Nothing fires on its own. Start every task here, including the ones you're sure about.
  2. Triggered and approvedSomething happens in the business, a draft appears in a queue, a person approves it. You've stopped deciding. You haven't stopped checking.
  3. RunningIt happens and nobody looks. The one everybody pictures when they say automation. Get here last, and only where three weeks of stage two gave you nothing to fix.

The ladder is useful because it tells you what to buy, and when.

Stage one needs no software at all. Stage two usually needs something small, and it's often a shared inbox or an approval queue you already pay for. Stage three is the only one worth a monthly bill, and by the time you get there you'll know exactly what you're buying, which is the whole reason for arriving in this order.

Figure 3 / Stack

Three stages you cannot swap, and only the top one is ever worth a monthly bill.

What automated looks like at eleven people

  1. Layer 3 of 3

    Running

    Nobody looks. The only stage worth a monthly bill.

  2. Layer 2 of 3

    Triggered and approved

    A person approves a queued draft. Usually needs something small.

  3. Layer 1 of 3

    Drafted

    A person reads it and sends it. Needs no software.

RunSimpler

The review step nobody builds

One tally mark. That's it. No dashboard, because somebody will actually do this one.

The mark isn't a verdict on the tool. It's a count of how often a person had to step in, which is the only number that tells you whether you can stop watching.

Build the review step this week
  • Put a sheet of paper next to whoever reads the output.
  • Make one mark every time a person has to fix something before it goes out.
  • Add the marks up on Friday and write the number at the bottom.
  • Stop and look if the number climbs two weeks running.
  • Take the next task off your scored list when it's flat and low.

Flat and low is the signal to move on. Climbing is the signal that something upstream changed, and it's usually the input rather than the tool.

Can you name the step that breaks?

Here's the question that decides whether any of this works.

Pick the task you just scored highest. Now say which step in it goes wrong most, and what it breaks further along when it does. If that's easy, you're ready and you don't need us. If that's hard to answer, automation isn't your problem yet.

The free AI readiness check is twenty questions, one tap each, about four minutes. Your finding and all five areas land on screen when you finish. Two of the five are whether the work is written down and whether anyone checks it. The score above depends on both.

Who actually needs help with this

Three routes. We teach your people to do this themselves, we build the thing and hand it over, or we run it for you every month.

The scoring exercise above is yours either way and it takes fifteen minutes. If you'd rather not guess at which route fits, the honest version of that question is a separate read. It includes when not to hire anybody.

Questions and answers