AI for Small Business: What It Does, Job by Job

Most writing about artificial intelligence and small firms is either a list of tools or a warning. This is neither. It is an attempt to go through the actual jobs in a business run by one or two people and say, for each one, whether handing it to a machine currently returns more than it costs. Some do. Several do not, and a few are actively dangerous to automate.

The framing matters because the honest answer is uneven. A tool that drafts a product description well can be catastrophic at quoting your returns policy, and the difference is not the tool. It is whether a wrong answer gets caught before anyone acts on it.

How many small businesses actually use AI?

Fewer than the marketing suggests, and the number has been flat for the smallest firms for over a year. The US Census Bureau's Business Trends and Outlook Survey, collecting from 14 December 2025 to 3 May 2026, found overall AI use hovering between 17 and 20 percent of businesses, with firms of fewer than 20 employees below 20 percent and not changing significantly across the period.

The size gradient is steep. Firms with at least 250 employees reported 37 percent use, and firms with 100 to 249 employees 32 percent. If you run a shop alone and feel behind, the data says you are not behind your peers. You are behind companies that employ someone whose job is to evaluate software.

Sector matters as much as size. The same survey put the information sector at 39.7 percent and finance and insurance at 33.9 percent, while retail trade sat near 14 percent with about 17 percent expected within six months. Retail is a laggard in these figures, which is either an argument that the tools do not fit retail yet or an argument that there is room to move early. Both readings are defensible.

Europe reads similarly. Eurostat found that 20.0 percent of EU enterprises with ten or more employees used AI in 2025, up from 13.5 percent the year before and 7.7 percent in 2021. The country spread is enormous, from Denmark at 42.0 percent down to Romania at 5.2 percent, which is a reminder that adoption tracks local business culture and available support more than it tracks the technology.

One more figure is worth holding onto, because it cuts against the usual narrative. The Federal Reserve's April 2026 note on measuring adoption records that around 41 percent of workers report using generative AI at work while only about 18 percent of firms report adopting it. Individual people are far ahead of their employers. In a one person business those two numbers are the same person, which is an advantage: you do not have to wait for anyone to approve anything.

What is AI actually used for, when businesses use it?

Reading, more than writing. Eurostat's breakdown for 2025 puts text analysis at the top with 11.8 percent of enterprises, ahead of image and video generation at 9.5 percent and text generation at 8.8 percent, with speech to text at 7.2 percent.

That ordering surprises people who assume the whole category is about producing content. Analysing written language grew fastest too, up 4.9 percentage points in a year. In a small business, reading looks like this: sorting an inbox, pulling the order number out of a message, deciding whether a review is a complaint or a question, summarising forty pieces of supplier correspondence into a decision.

The jobs, and whether handing them over pays

What follows is a judgement, not a measurement, and it is built on one test: when this goes wrong, who finds out and how much does it cost. A job where you catch the error in thirty seconds is a good candidate. A job where the customer catches it is not.

Writing: the job it does best and the one it fakes hardest

Drafting is where the return is most obvious. A model will produce a competent first version of a listing, a category page, an email or a supplier enquiry in seconds, and a competent first version is worth real time when the alternative is a blank page at eleven at night.

The failure is specific and it is always the same. Asked to describe a product it has never seen, a model fills gaps with plausible detail. It will give you a fabric weight and a country of origin that you never supplied. None of that is a lie in the model's terms, it is a guess presented in the same tone as a fact, and that tone is the problem. We wrote up a working method for this in our piece on writing product descriptions with AI without inventing the product, and the core of it is that you supply every fact and the model supplies only the sentences.

The other quiet failure is sameness. Run forty listings through the same prompt and you get forty listings with the same rhythm, which reads as a catalogue nobody wrote. Vary the instruction per category, or accept that your listings will sound like everyone else's.

Customer service: worth it in one direction only

Reading incoming messages is safe and valuable. Sending outgoing messages unsupervised is where small shops get hurt, because a chatbot that invents a returns window has made a promise a customer can reasonably hold you to.

The split that works for a very small business is to let the machine triage and draft, and to keep your finger on send. Triage alone is worth the subscription: knowing which six of forty messages need you today is most of the value, and getting it wrong costs nothing because nothing was deleted. We went through how to sequence that in which support tickets to give AI first.

Numbers: good at describing, poor at deciding

Asking a model to explain what happened in your sales last month works well, because the data is in front of it and the answer is checkable. Asking it to decide what to order next month works badly, because the thing that determines the answer is usually not in the data at all.

The Federal Reserve note is careful on exactly this point, declining to make productivity claims and flagging that the differences between surveys come down to question framing and what counts as material use. That caution is appropriate for a solo business too. A tool that tells you your Tuesday afternoons are dead is useful. A tool that tells you to buy 400 units is making a bet with your money using less context than you have.

Images: the fastest win and the easiest way to get delisted

Cleaning up a background, matching lighting across a catalogue, or generating a lifestyle scene are all fast and cheap now, and the quality is good enough for most storefronts. Eurostat's 9.5 percent for image and video generation makes this the second most adopted use in Europe.

The constraint is not quality, it is honesty about the product. Every major marketplace treats an image that misrepresents what arrives as a listing violation, and a generated image that quietly changes a colour or a texture does exactly that. Edit the scene, never the item.

What must you tell customers, legally?

If you serve EU customers and you run a chatbot, you must tell people they are talking to a machine, and you must do it at the start of the first interaction. That obligation has been live since 2 August 2026 under Article 50 of the EU AI Act.

The European Commission's own guidance is unusually concrete about form. The notice must be clear and distinguishable, and perceivable without any specific technical tools, which rules out burying it in terms and conditions. There is a sensible carve out: the duty does not apply where it is obvious the person is dealing with an AI system.

Generated text gets a narrower rule than most coverage implies. The labelling duty for AI generated text attaches to material published to inform the public on matters of public interest, and text that went through genuine human review and editorial control does not need a label. A product description you wrote with a model and then checked is not a regulated disclosure. A synthetic image of a person in your advertising is closer to the line.

We keep a fuller map of which obligations land when in the EU AI Act tiers and timeline, and the practical chatbot wording in what your chatbot has to disclose.

What should you refuse to buy?

Anything sold on an earnings promise. The Federal Trade Commission's sweep against AI marketing claims is the clearest available guide to what predatory looks like in this market, and the pattern repeats with unusual consistency.

In that action the FTC alleged consumer losses of at least 25 million dollars against one operation selling AI powered online storefronts, and more than 15.9 million dollars against another promising guaranteed income from the same idea. A third charged up to 35,000 dollars a customer while advertising 10,000 dollars a month in earnings. The Commission's summary of the principle is worth keeping: there is no AI exemption from the laws on the books.

Using AI tools to trick, mislead, or defraud people is illegal.Lina M. Khan, FTC Chair

The same enforcement action included a company whose product generated fake consumer reviews at scale. If a tool offers to write your reviews, the tool is the liability, not the shortcut.

Two softer rules follow from the same instinct. Be suspicious of anything priced per seat when you are one person, and be suspicious of anything that requires your entire catalogue before it will show you a single output. Our breakdown of what small business owners actually pay for AI goes through the pricing shapes and which ones punish a small operator.

Where the money goes, and how much you need

The realistic floor for a very small business is one general assistant subscription, and for many owners that is the whole stack. Specialist tools earn their place only when a job is both frequent and painful, and frequency is the part people misjudge. A task you do twice a year does not justify a monthly fee, however annoying it is.

A useful test before any purchase: write down how many times you did the job last month and how long it took. If the answer is under an hour in total, the tool cannot save you an hour, and no amount of capability changes that arithmetic. This sounds obvious written down and it is the single most common reason small businesses end up with six subscriptions and no time back.

Watch for the second cost too, which is the review time you now spend checking output. A tool that drafts in ten seconds and takes four minutes to verify has not saved four minutes, it has moved them. That is still often a good trade, because reviewing is easier than composing, but it is not the trade the marketing describes.

A sequence that works

Pick the job you do most often that no customer ever sees. For most shops that is inbox triage, or turning a supplier spreadsheet into something usable. Do it manually one more time, write down the steps, then hand exactly those steps over and compare the output with what you would have produced.

Run that comparison for two weeks before you trust it. Two weeks is long enough to meet the edge cases, which is where every one of these tools falls over, and short enough that you have not built a process around something that does not work.

Only then move outward, one job at a time, and keep the rule that anything reaching a customer unreviewed needs a much higher bar than anything reaching only you. If you are building the storefront itself rather than bolting tools onto a rented one, the same principle applies to the code: our AI store builder generates the shop and hands you the repository, because a system you can read is a system you can correct.

Questions small owners actually ask

Is AI for small business different from AI for a big company?

Yes, in one structural way. A large company buys AI to remove a bottleneck between departments, while a solo owner buys it to remove a bottleneck between two of their own tasks. That changes what is worth paying for: coordination features are worthless to you, and speed on a single repeated job is everything.

It also changes the risk. In a large firm a bad output passes through someone else before it reaches a customer. In yours it does not, unless you deliberately build that step in. Most of the discipline described above exists to reconstruct a review layer that a bigger organisation gets for free.

Which AI tools for small business are worth a subscription?

One general assistant, and after that only tools that attach to a job you did more than ten times last month. That threshold is arbitrary and it is close enough, because below it the time saved cannot exceed the time spent learning the tool and checking its work.

Rank the candidates by how easily you can verify the output. AI product descriptions are easy to verify because you own the product and can read the draft against it. AI bookkeeping is harder, because a miscategorised expense looks correct until a quarter has passed. AI customer service sits in between, safe when it drafts and risky when it sends.

Do I have to tell people when I use AI?

For a chatbot serving EU customers, yes, and the chatbot disclosure has to appear at the start of the conversation rather than in a policy page. For text you wrote with a model and then edited yourself, generally no, because human editorial control is the condition the rule turns on.

Outside the EU the obligation is usually indirect but no weaker in practice. Advertising law already forbids misleading claims, and a machine wrote it is not a defence when the claim turns out to be false. Platform rules add their own layer, and marketplaces have moved faster than legislators on labelling generated imagery.

What happens if I do nothing this year?

Very little, immediately, which is why doing nothing is such a comfortable choice. The Census numbers show most of your direct peers have also done nothing, so there is no competitive cliff waiting in the next few months.

The cost shows up later and sideways. It arrives as the hours you did not get back, and as a slower response time than a competitor who triages their inbox with a machine. Neither is dramatic. Both compound.

What this looks like in a year

Two things in the data point the same way. Adoption among the smallest firms has been flat while adoption among individual workers has not, and the gap between 18 percent of firms and 41 percent of workers is a gap that closes from below. The tools are already in your hands whether or not the business has a policy about them.

The second is that retail sits below the cross sector average in the Census figures, at around 14 percent against 17 to 20 percent overall. Sectors that lag usually lag because the obvious applications do not fit, and then they catch up quickly when someone finds the ones that do. Reading, rather than generating, is the likeliest candidate, because a shop drowns in text long before it runs short of it.

The advice does not change with the tools. Hand over the jobs where a mistake is cheap, keep the ones where it is not, and write down which is which before somebody sells you a subscription that decides for you.