AI for Small Business: What the Surveys Still Say

The useful question for a small business is which job actually changes. Two surveys from 2025 are still the clean public counts. They do not measure the same thing. One asks whether a small business uses generative AI. The other asks whether a business used AI to make its goods or deliver its service. You can say yes to the first and no to the second, and be honest both times. If you average them into one “how far behind are we” number, you may buy the wrong project.

Cited in this note

58%

We wrote out the full split in the help note. If you want to know what to do about the gap, read the strategy note. If you want to score one firm instead of a whole country, use the assessment. A survey counts a crowd. Your shop is one set of hands, one set of files, and one job that is slow this week. Start with that job.

People ask us which percentage to believe. We do not treat either one as a verdict on a single owner. What we trust is the wording of the question. If you skip the wording, you can spend a whole season chasing a rank that was never about your counter.

Two questions, two answers

The Chamber’s 2025 report says 58% of small businesses use generative AI, up from 40% in 2024 and 23% in 2023.¹ That question is about using a tool to write a product note, draft a reply, or ask for a checklist. It counts the owner who opened the tool on a Tuesday and got something good enough to edit. It does not ask whether a model made the goods on the shelf. It does not ask whether a model did the work the customer paid for.

Statistics Canada found that 12.2% of businesses used AI to make goods or deliver services in 2025, up from 6.1% a year earlier.² That question is harder and narrower, because it is about the thing you sell. A bakery that uses a tool to write the weekly email can sit inside the Chamber number and outside the Statistics Canada number. A shop where a model suggests orders and a person places them is close to the second question. It may still not meet it. Read the verbs, which are make and deliver. They are not draft, brainstorm, or “we tried it once.”

A firm can say yes to the Chamber question and no to the Statistics Canada question. Plan from the job you can name, and do not average the two percentages into one readiness rate. This site does not publish a combined score from the Readiness Check. If someone hands you one blended number made from these two, they have glued different questions together.

A worked example

Walk through a home-goods shop with us. This is an example and not a study, and we are not giving it a score. The owner buys from a few makers and from one wholesaler. A floor lead works the counter and knows which lamp has a chip in the glaze. A bookkeeper comes in part of the week and dislikes surprises in her inbox. The shop has no IT person, and the team talks by group text.

Last winter the owner started using a generative tool for product descriptions. She pastes in a few notes from the maker and asks for a short paragraph in her own voice. Then she edits out anything that sounds like a catalog from nowhere. The floor lead does the same for a social post when a new table arrives. He deletes the line that says the wood is oak when it is not. That is real use, so on the Chamber question this shop says yes.

Nothing about that use makes the lamp or delivers the sale. The owner still picks the maker. The floor lead still wraps the lamp and tells the truth about the chip. He also still calls the customer when the freight is late. The bookkeeper still matches each invoice to the packing slip with her own eyes. On the Statistics Canada question, this shop says no. AI is not making the goods, and it is not doing the work for the customer. Both answers are correct. A consultant who says “you are an AI business now” or “you are behind” has not listened.

Here is a job that might move the shop toward the second question, if they ever want that. The wholesaler’s invoices arrive as scans, and the bookkeeper retypes every line. A tool could read each scan and suggest the lines while she still posts them herself. That is closer to the back-office service than a nicer lamp description is. It is still not the shop “running on AI.” It is one job with one checker, and it touches work they already do to get paid. We would talk about that job for an hour before we ever mentioned the Chamber number.

The trap in this shop is pride that points the wrong way. The owner feels modern because her descriptions got better. Her nephew told her that most small businesses already do this. He was pointing at the Chamber number without saying what the question was. She also feels secretly behind, because a podcast said only a small slice of firms use AI in production. That podcast was pointing at a number like the Statistics Canada one, again without the question. She is not a laggard and she is not finished. She has a drafting habit and an untouched pile of invoices.

How we’d run it

We do not start by asking an owner where they rank. We start by pulling the two questions apart, and then we pick one job.

First, we ask what they already open in a normal week. The shop owner will name the writing tool. We write that down as use in the Chamber sense. We do not praise it as a plan, and we do not scold it. A product note that a person edits is a fair use of an afternoon. It is not the same as AI doing the work for the customer.

Second, we ask what the customer pays for and what has to be true for the shop to earn it. In this shop the customer pays for the right object, in the shape the floor lead described. It must be handed over or shipped, with books that explain the invoice. We look for any step in that chain where a model already acts. Usually there is none, and we say so. The Statistics Canada answer stays a no. That is useful, because it stops the owner from bragging about a skill she does not have.

Third, we pick one job that could change the chain. We describe it in a sentence a new hire could follow. For this shop we like the scanned invoice better than the caption. The sentence is to read the wholesaler’s scan, suggest the lines, and wait for the bookkeeper to post them. If the owner would rather automate the caption, we will do it. But first she must admit that it does not change delivery. That can still be worth a small habit, but it is not worth a project with a name.

Fourth, we name the checker and the failure. The bookkeeper is the checker. A failure is a line that does not match the packing slip, or a vendor she does not know. If she cannot see the original scan next to the suggestion, the job is not safe. We do not turn it on. This is the same care we give any other install, and a survey cannot do it for her.

Fifth, we refuse the blended score. We will not tell this owner that she sits at some midpoint between the two numbers. Those groups and those verbs are different. The assessment scores her files, her people, and her job. The strategy note is where the one-page plan goes once the job is chosen. The help note is the longer story of the split, for the day a partner still wants to argue from a headline.

Common mistakes

Quoting one percentage in a staff meeting is the fastest way to start a useless argument. When someone says “most small businesses already use it,” they are pointing at the Chamber’s 58% in 2025. That was up from 40% in 2024 and 23% in 2023. A person on the floor hears that they are late. They are not late at wrapping lamps, because they are being compared on a question about generative tools. Say the question out loud or leave the number out.

The mirror mistake is shame from the smaller number. In the Statistics Canada count, 12.2% of businesses used AI to make goods or deliver services in 2025. That was up from 6.1% a year earlier. An owner hears “almost nobody is really doing this” and either freezes or decides the whole topic is hype. Neither reaction picks a job. The smaller number describes a strict question and says nothing about your character.

Averaging the two rates is the mistake we will interrupt every time. A midpoint is not a readiness rate, because the firms, the countries, and the verbs all differ. This site does not publish a combined score from the Readiness Check. We do not want a homemade one in a spreadsheet either. If you want a number for your own business, it has to be about your own business. A national blend cannot tell the bookkeeper whether a scan is readable.

Buying a tool to close the feeling gap is the costly version. The owner sees the gap between “we write notes” and “AI runs the shop.” So she buys something that promises to run the shop. Neither survey asked her to do that. The surveys never named her wholesaler, her chipped glaze, or her bookkeeper’s part-time days. A purchase that ignores those facts will stall for the usual reasons, and then people will blame the survey.

Treating a yes on the Chamber question as a marketing claim is how trust frays. If a sign in the window says “we use AI” and you only draft captions, a customer will expect something you do not do. We would rather the shop say nothing. The floor lead’s honesty about the glaze is worth more than a badge.

What we’d do

We would plan from the job. We would say the Statistics Canada question out loud before the Chamber one. That is a preference and not a law. Drafting help is legitimate, but it is the easier yes, and it flatters the owner. The question about making goods or doing the service is the one that changes pay, mistakes, and what the customer gets. If the owner only wants drafting help, we will help her keep it small and checked by a person. We will not call it a change in how the shop runs.

We would not tell the home-goods owner that she is behind the 58%. We would not tell her she has “arrived” because she might be inside it either. Arriving means you can name a job on a slow Tuesday and a person checks it. The climb from 23% in 2023 to 40% in 2024 to 58% in 2025 shows chat tool use spreading. The climb from 6.1% to 12.2% in the Statistics Canada numbers shows a stricter habit spreading more quietly. Both can be true, and your shop does not have to pretend to be either crowd.

If a partner brings you a chart this week, ask which question was on the survey. If they cannot say, put the chart down and open the assessment. If they can say, and the question is about use and not delivery, read the help note before you spend money. When you are ready to choose the job, the strategy note is the page we would write with you. We would rather be plain and specific than armed with a number about someone else’s work.