AI automation for small businesses: where to start

People now talk about artificial intelligence in business the way they talked about going digital ten years ago: a lot, and almost always without saying anything concrete. Revolution, transformation, new paradigm. Then you walk into a real company and you find a person at seven in the evening copying figures from an email into a spreadsheet.

This article tries to stay on that person’s side. No scenarios, no future: four processes with names, three things I actually built, and an honest answer about cost.

What I mean by AI automation

An automation is a piece of work a person does today and a machine does from tomorrow, with the person checking instead of doing. So far there is nothing new: we have been automating things for decades.

What changes with artificial intelligence is the kind of work you can move across. Before, you could only automate tasks with rigid rules: if this arrives, do that. Anything that meant understanding a text, summarising a document, writing a sensible reply or making sense of a badly written message had to stay with a person. Now part of that can move.

The important bit, and the one usually skipped: AI is not as precise as a traditional piece of software. A calculation comes out the same every time, an interpretation does not. So you use it where the odd mistake is manageable and where somebody looks at the result before it goes out, and you keep it away from anything that needs to be exactly right. Anyone telling you it is infallible either has never used it seriously, or has something to sell you.

The 4 processes to start from

In small and medium companies the candidates are almost always these four.

Answering customers

The questions that come in every day are always the same five: how much does it cost, how long does it take, are you open, where are you, do you also do this. An assistant on WhatsApp or on the site can answer those and pass you only the real enquiries. It does not take the relationship away from you: it takes away the time spent writing the same sentence for the thousandth time.

Quotes and documents

If your quotes change only in the numbers and the line items, rewriting them by hand is work nobody pays you for. Same for job sheets, forms, cover letters, documents to be filled in from data that already sits somewhere. This is the process where the saving shows up first.

Orders and kitchen tickets

Whenever a piece of information has to reach several people at the same moment and today travels on a sheet of paper or a voice message, there is almost always something to fix. It is not even the most “intelligent” part: often it is just a matter of showing the same thing to everybody at once.

Reminders and follow ups

The deadlines, the renewals, the customers who asked for a quote and never came back. These are the things you forget when the day goes sideways, and they are also the ones worth the most: a customer contacted again at the right moment is worth more than ten new leads.

All four are explained at length on the AI automation for businesses page.

Three real examples

I built these myself. And before building them for anybody else I built them for me: I run my own company with a team of AI agents I put together one piece at a time, so what I am offering you has already been tested in my own house. I do not give client names, but the processes are exactly these.

A sales agent who was making his content by hand. He wanted to be visible online and he was building every piece one bit at a time: writing the text, editing the video, adding subtitles and his logo. One at a time, almost always in the evening, aiming for around twenty pieces a month: impossible at that speed. Now he sends a message to a bot while he is driving, even just a voice note, and gets back video and written posts ready to check. Publishing stays his decision: the system posts nothing on its own.

A takeaway kitchen where the order travelled on a slip of paper. The customer fills in a slip, takes it to the till, and that slip then has to physically go round every station in the kitchen. While it sits at one station the others do not know they should be starting: at peak time that means dishes coming out out of step, and queues. I built a system where the order appears on every screen at the same moment, with the number large, and the collection counter can see at a glance which stations are done and which are not.

A health practice with paper records. Intake forms and patient records on paper, bookings scattered between a platform, WhatsApp and the phone. I built a system where the intake is guided and the patient can fill it in before arriving, where certain warning signs trigger an automatic alert, and where exercise sheets are generated and sent as PDFs instead of being rewritten every time.

An honest note: in two of these three cases the artificial intelligence part is small. I include them anyway, because the point is not to use AI. The point is to remove pointless work, and sometimes the right tool is simpler and costs less. If your problem can be solved without AI, I will tell you: what you need might be custom software built around your work.

Where you actually start

Not with the digital transformation of the company. With one process.

The method that works is dull and it is this. Take the thing you redo every week that weighs on you most, not the most complicated one: the most boring one. Look at how much time it really takes over a month, by counting, not by guessing. Then automate that one and only that one, and let it run for a few weeks with somebody checking the results.

If it works, you take another one. If it does not work you have lost a small project and you have not stopped the company: that is the whole difference between starting small and starting big.

The checking part is worth two words of its own, because it is what decides whether an automation lasts. On one job I did, short voice messages were being misheard by the system, and misheard differently every time: the same two second clip came back transcribed four different ways, all of them wrong. The fix was not to hope it would improve: the system now reads those messages twice and, if the two readings do not match, instead of guessing it drops them and asks for the message again. That is how these things get built: knowing they can be wrong, and deciding in advance what happens when they are.

What does it cost?

It depends on the process, and on this page you will not find a price list, for the same reason you never find one when it comes to software built to measure: an assistant that answers five questions and a chain that prepares documents by pulling from three different systems are jobs of wildly different size.

What I can tell you is how I think about it. I price on value, not on hours: how much time that thing gives you back, every month, and what that time is worth to you. If the sum does not come out in your favour, I say so and we do not do it. I have no interest in building you something you use three times and abandon.

If you want to see how I think about prices in general, I wrote it out in full when talking about how much a website costs: the method is the same.

If you want to work out where you would start

You do not need to know already what to automate. Quite the opposite: almost nobody does, and the piece that can be automated is often the one you had not thought of.

Tell me what your day looks like and which two or three things you keep doing the same way over and over. I will tell you whether they can be taken off your hands, how, and whether it is worth it. If it is not worth it, that is as good an answer as any other.