You have been told, repeatedly and by people who seem confident, that AI is transforming small businesses. You have possibly already tried something: signed up, spent an evening on the setup, used it enthusiastically for a fortnight, and then quietly stopped. The subscription is probably still going out.
That experience is close to universal, and it is not because you did it wrong. The useful question is not whether AI is real; it is what actually works in a business the size of yours. AI for a small business works reliably on a narrow set of jobs, fails predictably on another set, and gets abandoned most often when it is used on the right job in the wrong way. Surveys through 2026 keep finding the same gap: most small firms say they use AI somewhere, and far fewer have it genuinely running inside a process without someone watching it. This is a guide to which side of that gap each idea belongs on.
Why the tool you bought is gathering dust
Before the lists, the failure pattern, because it explains both of them. In our experience the abandoned tool almost always failed for one of five reasons, and only one of them is about the technology being bad.
- The setup was brutal. It assumed you had your information in a form you do not have: a clean product list, documented policies, a tidy customer database. The demo used the vendor's sample data. Yours is in six places and three of them disagree.
- It needed babysitting. It produced something plausible every time, including when it was wrong, so checking its work cost as much as doing the work. Anything that always answers and never says "I am not sure" creates supervision rather than removing it.
- It solved a problem you did not have. The demo was impressive, the job it did was not one that was actually costing you money, and impressive wears off in about three weeks.
- Nobody owned it. It was somebody's enthusiasm, that person got busy, and there was no process that would break if it stopped. Anything optional in a small business is eventually optional.
- The process underneath was broken. This is the big one. Automating a mess produces a faster mess, and now the mess is harder to see. If quotes go out late because nobody decides the price, no tool fixes that; it just sends the indecision quicker.
AI for a small business: what actually works, in five jobs
Five jobs. What they have in common is worth noticing: high volume, a clear right answer that exists somewhere in writing, and a cheap failure. When those three are true, this technology is genuinely very good.
| The job | What it looks like in practice | Why it works | Typical monthly cost |
|---|---|---|---|
| Answering the same questions | The forty questions customers ask before buying, answered instantly from your own policies and product information, at eleven at night and on Sundays, with anything unusual passed to a person. | The answers already exist in writing, the volume is high, and being unsure is safe as long as it hands over. | $100 to $600 |
| Turning mess into fields | Supplier invoices, delivery notes, timesheets, application forms and emailed orders read and turned into rows in your system, with anything ambiguous flagged rather than guessed. | The input is unstructured and the output is checkable in seconds, so mistakes are caught by the next step rather than shipped. | $80 to $500 |
| First drafts nobody enjoys writing | Product descriptions, listing copy, standard replies, job adverts, meeting notes turned into actions. A person edits every one before it leaves. | A mediocre draft in ten seconds beats a blank page in ten minutes, and the human edit catches what matters. | $20 to $200 |
| Summarising and routing | A long email thread, a recorded call or a support queue turned into who needs to do what, sorted so the urgent ones surface. | Nothing is decided automatically. It changes the order you see things in, which is where most of the time actually goes. | $50 to $400 |
| Finding things in your own documents | Asking a question and getting the right paragraph of the right manual, contract, price list or policy, with a link to the source. | The answer is verifiable in one click, so the cost of being wrong is a second of your time. | $100 to $700 |
Costs cover the tooling for a small business at modest volume and exclude the work of setting it up. Every one of these is a candidate for an off-the-shelf product before anything custom.
What does not work yet, whatever the demonstration showed
- Anything where being wrong is expensive and hard to spot. Quoting a price, giving tax or legal or medical guidance, telling a customer their warranty position. The failure mode is a confident, plausible, wrong answer that nobody catches until it has been acted on. Use it to draft for a human, never to decide.
- Long chains of unsupervised steps. An assistant that reads the email, updates the system, orders the part and tells the customer sounds like the future and behaves like a rumour: each step is mostly right, and mostly right five times in a row is often wrong. Chain two steps with a human checkpoint, and revisit in a year.
- Replacing a person. What it replaces is a task. Businesses that budget for a headcount saving are consistently disappointed; the ones that get value buy back hours inside jobs people already do. If a supplier's proposal only works if someone leaves, be sceptical of the whole proposal.
- Fixing a process nobody has defined. If two people in your business would answer a customer question differently, an assistant cannot resolve that. It will pick one, confidently, forever. Decide the answer first; that is why our workflow automation work starts with the process rather than the tool.
- Anything that depends on knowledge you never wrote down. The reason you always give that customer a discount, the supplier you never use in December, which jobs are worth doing. It is not in the system, so it cannot be in the assistant, and the output will be subtly wrong in ways only you can see.
Buy it, or have it built?
Almost always buy first, and be suspicious of anyone who tells you otherwise before they have asked what you do. A subscription costs tens of dollars a month, works this afternoon, and tells you within a fortnight whether the job was worth automating at all. That answer is worth more than the money you would save building it.
Building earns its place in four specific situations: when it has to be grounded in your own documents and product data rather than general knowledge, when it has to reach into systems that have no off-the-shelf connector, when your rules genuinely differ from everyone else's, or when the per-seat and per-resolution charges at your volume have grown past the cost of running your own. Notice that three of the four are things you can only discover after buying something first.
Where to start, in order, for almost nothing
- Write down last week honestly. Not what feels slow: what actually took hours. Most owners are wrong about this until they check, and the winner is usually something dull and repetitive that nobody thinks of as a problem.
- Pick the one with the highest volume and the cheapest mistake. Not the most impressive. Volume is what pays for it and a cheap mistake is what lets you deploy it without supervising it.
- Fix the process on paper first. If a person could not follow your written instructions for this task, no assistant can either. This step is free and it is the one that gets skipped.
- Buy something for a month. Use it properly, every day, for four weeks. Decide with your own data, not with a demonstration.
- Design the escalation before you launch it. What must always reach a person? Anything about money, complaints, health, legal position, or anything the assistant is not confident about. Write the rule down and test it deliberately with questions your content cannot answer.
- Measure the before. Count the hours or the response times for a week beforehand. Without that number you will be arguing about a feeling in three months.
- Then, if it earned its place, consider building. Now you know the volume, the edge cases and the value, which is exactly what makes a custom build cheap to scope and safe to commission.
If you get to step four and the honest answer is that the tool did not save anything, you have spent about fifty dollars to avoid a project. That is the best possible outcome of this exercise and it happens often.
What it costs, realistically
For a small business doing this sensibly, illustratively $50 to $300 a month in tooling for one or two of the jobs above, plus the hours of whoever owns it. A custom assistant grounded in your own content is a different order of thing: illustratively $8,000 to $60,000 to build and $100 to $2,000 a month to run, which our guide to what an AI assistant costs sets out properly, including why the monthly figure is a cost per conversation rather than a licence fee.
The cost that nobody budgets for is attention. Every tool needs an owner, a review of what it got wrong, and a decision each quarter about whether it still earns its subscription. Three tools with an owner beat ten without one, comfortably.
Not for you if
The decision you are actually making
You are not deciding whether to adopt AI. You are deciding which one repetitive job in your business is worth removing from a human this quarter, and whether the answer to that job already exists somewhere in writing. Where it does, this technology is unexciting and genuinely useful, which is the highest compliment available in business software. Where it does not, no amount of spending will conjure it, and the honest first step is writing the answer down yourself.