"What should I automate first in my small business?" Most owners answer with whatever annoys them most this week. That is a fair place to start, but it is a poor way to finish, because the task that annoys you most is often the one with the messiest rules and the costliest mistakes. This article is for the owner of a company with ten to two hundred people who wants a way to choose. It explains the two kinds of automation, runs through five common candidates (invoices and data entry, scheduling, order status, reporting and onboarding paperwork), gives a five-question score you can work out on a sheet of paper, lists what not to automate first, and ends with a two-week trial. It is about automation choices in general. If you already know you want to start with AI and need to pick the project, picking your first AI project is the next step.

I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them, so I have seen how many back-office tasks a growing company carries that nobody ever chose to keep. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs. This is not legal advice. Where the law could apply to your company, talk to your own counsel.

Two kinds of automation, and why it matters

Rules-based automation follows a fixed rule: when this happens, do that, the same way every time. A form comes in and a confirmation goes out. Stock drops below a set level and the warehouse is told. An invoice over a set amount goes to a named person for approval. No judgment is involved, nothing is interpreted, and it does not get confused. It is usually the cheapest and most predictable option, and it was around long before anyone talked about AI.

AI is for the steps where a fixed rule cannot cope, usually because the input is messy: a customer email written in a hurry, a supplier invoice in a layout you have never seen, a long note that needs a summary. The tool reads the mess and produces a draft or a best guess, and a person normally checks it before it matters.

Many good first projects use both. A rule moves the work along; AI handles one messy step inside it; a person approves the result. The point for choosing is simple: if the rules are fully known, you may not need AI at all, and paying for it adds cost and uncertainty for no gain. If the rules are fuzzy, AI may help, with a person checking.

Five common candidates

Here is how each one tends to look. Your own business will differ, so treat these as examples to score, not a ranking.

Invoices and data entry. Typing figures from supplier invoices, orders or forms into your accounting or order system is frequent, dull and easy to get wrong when tired. If the invoices arrive in a handful of fixed layouts, a rule can often do most of it. If they arrive in dozens of layouts, AI can help pull out the fields, and a person checks the totals before anything is paid. Keep a person on anything that moves money. Can AI do bookkeeping for a small business covers where it helps and where it does not.

Scheduling. Booking jobs, appointments, deliveries or shifts follows rules about availability, travel and length of job. Rules-based tools do this well when the constraints are clear. Where customers ask in free text, such as "sometime after school next week," AI can turn the request into a proposed slot for a person to confirm.

Order status. "Where is my order?" is a question customers ask all day, and the answer sits in a system you already have. A rule that sends a status message when a stage changes removes many of those questions before they arrive. For the questions that still come in, AI can draft an answer from the record for a person to send. Can AI answer customer emails for a small business shows how that works.

Reporting. The weekly or monthly report that somebody assembles by copying figures from three places is a strong candidate when the sources are stable. A rule can collect the figures into the same layout each time. AI can add a short written summary of what changed, which the owner then checks against the figures. The risk is trusting a neat summary of wrong numbers, so someone must own the source data.

Onboarding paperwork. New-hire forms, equipment requests and welcome messages follow the same steps for each person. A rule can send the right forms in the right order and remind people who have not returned them. Keep contracts, pay terms and anything with legal effect with your own advisers and a person who signs off.

The paper score: five questions

Get a sheet of paper. Write each candidate task on its own line. For each, score five questions from one (weak) to three (strong). Write the scores as words if you like, or tick marks; what matters is the total.

Frequency. How often does the task happen? One for rarely, two for weekly, three for daily or many times a day. A task done twice a year is almost never worth automating first.

Time. How much total human time does it eat? One for a few minutes now and then, two for an hour or two a week, three for much of somebody's week. Ask the person who does it; owners guess low.

Error cost. How costly is a mistake? Here a high score means a mistake is cheap and easy to catch. Three if a wrong result is small and a person would notice it. Two if it costs some money or an awkward email. One if it touches payments, contracts, customer promises or anything you could not undo.

Rules clarity. Can you write down the steps so a new hire could follow them? Three if yes, the same way every time. Two if mostly, with a few judgment calls. One if the answer is "it depends" and the person who knows is the only one who can say.

Data readiness. Is the information already in one place, in a usable form? Three if it sits in a system or a tidy spreadsheet. Two if it is in several places but findable. One if it is on paper, in people's heads or in a folder nobody has tidied.

Add the five scores. A total of twelve or more is a strong first candidate. Nine to eleven is a possible second project. Below nine, leave it for now. Then apply two overrides. If the error-cost score is one, do not make it your first, whatever the total. If the data-readiness score is one, fix the data first, because no tool works well on information nobody can find.

Five minutes on this sheet, done with the person who actually does each task, usually changes the answer. The winner is often a task you would not have guessed, because it is dull enough that nobody had complained about it.

What not to automate first

For a first project, leave these alone, and keep them with a named person:

None of these is off limits forever. They suit a second or third project, after you have seen how the first one behaved. If you are wondering how far this can go, can AI run my business sets out what a tool can take over and what stays with you.

A two-week trial

Run the winner as a trial, not a launch.

Before you start, write down the number. For a few days, note how long the task takes now, how often it is done and how many mistakes come back. Your memory of how long it takes is unreliable, and you will want the real figure later.

Week one: run it beside the old way. The new method does the task and a person does it the old way as well, or checks every single result. Note differences. Do not switch anything off.

Week two: let it carry the work, with a person checking. The person reviews every result at the point where a mistake would cost something, and writes down what they changed.

At the end, compare. Hours saved, mistakes found, and how long the checking took. If checking took as long as doing it by hand, say so and try a different task. If it worked, widen slowly, and keep a person at the point of risk. For the cost side, how much does AI cost for a small business breaks the spend into plain parts, and if nobody on staff is technical, AI for a small business with no IT department covers running all this without one.

Be honest about the result. The aim is to learn where automation helps in your business, not to prove that it does.

Picture a plumbing-supply wholesaler of sixty people. The owner lists five tasks: typing supplier invoices, booking deliveries, answering order-status calls, building the weekly sales report and sending new-starter forms. She sits down with the person who does each and scores them. Typing invoices scores low on error cost, because it moves money, and the supplier layouts vary. Delivery booking scores middling on rules clarity, because the dispatcher decides by feel. Order-status calls score high on frequency, time, data readiness and a cheap, catchable error, so she picks them. A rule sends a status message when an order ships, and a tool drafts replies to the remaining questions for a person to check. She writes down how long calls took before. After two weeks she compares. Nobody has been sold a platform, and she knows which task to try next. This is a hypothesis, a composite and not a client.

Mistakes that repeat

Choosing the task that annoys you most instead of scoring. Automating a process nobody has written down. Paying for AI where a fixed rule would do. Starting with anything that moves money or makes a promise. Skipping the number from before. Switching off the old way on day one. Not asking the person who does the task. Skipping your own counsel where the law could apply.

Frequently asked questions

What should I automate first in my small business?

The task that scores best on five questions: it happens often, takes real time, has a cheap and catchable mistake, follows rules you can write down, and uses data that is already in one place. Score each candidate on paper with the person who does it.

What is the difference between automation and AI?

Automation follows a fixed rule the same way every time. AI handles messy input, such as free-text emails or varied invoices, and produces a draft or a best guess for a person to check. Many good projects use a rule to move the work and AI for one messy step.

Do I need AI to automate a task?

Not always. If the rules are fully known, a fixed rule is cheaper and more predictable. Add AI only where the input is messy and a rule cannot cope.

What should a small business not automate first?

Prices, delivery promises, refunds, anything that moves money unchecked, pay and hiring decisions, legal and contract wording, broken processes and any customer-facing step with no person checking.

How long should an automation trial last?

Two weeks is enough to begin: one week beside the old method, one week carrying the work with a person checking, with the number from before written down and compared at the end.

Which tasks do small businesses usually automate?

Invoices and data entry, scheduling, order status messages, routine reports and onboarding paperwork are common. Score your own versions, because the rules and risks differ from company to company.

Is there anything the law requires me to check?

I cannot give legal advice, and it depends on your location, your industry and your contracts. Ask your own counsel what applies to your company, especially for anything touching employee or customer records.

Not sure where you stand? Take the 3-minute AI readiness scorecard.