How Can AI Help a Small Business? What Surveys Count

AI helps a small business when it takes over a named job that a person already does. It does not help by moving you closer to a headline. Two surveys from 2025 count AI in small firms, and they count different kinds of work. The U.S. Chamber asked who uses generative AI. Statistics Canada asked who used AI to make goods or to deliver services. A firm can say yes to the first and no to the second, and both answers can be honest. This note keeps those two claims apart, and then it names the jobs that owners hand over. The pressure guide asks the matching question for one business, and we write these notes on our site.

Cited in this note

58%

We sit with owners who have already been told they are late. The number they heard was usually the bigger one. So the first thing we do is find out what that number really measures. Help, for a shop or a clinic or a trades desk, looks like a draft that someone checks. It looks like a sort that someone trusts, or a reminder that someone still sends. It is not a rank on a national chart.

Generative use is not production use

The Chamber’s small business report says 58% of small businesses called themselves users of generative AI.¹ That is up from 40% in 2024 and 23% in 2023. That question is about generative AI. It does not ask whether AI makes the goods or the service. The owner said so, and no one watched the work to check. It also does not mean the tool is part of the product. An owner who drafts a customer email and then rewrites half of it can answer yes. We would answer yes with them. That yes is real, but it is a different fact from a system that makes the thing the customer bought.

Statistics Canada ran its Canadian Survey on Business Conditions for the second quarter of 2025. It found that 12.2% of businesses used AI to make goods or deliver services.² That is up from 6.1% in the second quarter of 2024. The release is dated 16 June 2025, and the survey ran from 1 April to 5 May 2025. The agency got 21,357 responses, and 9,103 of them came from businesses that used AI. Read those counts as a picture of who was asked, and not as a target. A big response still describes the questions that were asked. It does not turn the generative question into the production question.

Among the businesses that used AI, 35.7% used text analytics and 23.1% used marketing automation. Then 40.1% had made new workflows, and 38.9% had trained their staff. Those shares sit inside the group that already uses AI. They are not shares of every firm in the country, and they are not shares of the Chamber’s question. Text analytics means reading and sorting words. Marketing automation means sending messages and counting what comes back. A new workflow is a path that was not written down before. Training means a person was shown that path. Put that way, the list is practical, and it is also modest. None of the four says the business now runs itself.

Jobs an owner can hand over

The Chamber number describes a tool that someone opens. The StatCan number describes AI inside the making of the product or the service. A small business can do the first for a long time without ever doing the second. We think that is a sane order. Open the tool on a job you can check. Ask the production question later, when the goods or the service truly depend on the system. Most owners we sit with are nowhere near that later question, and they do not need to pretend they are.

The jobs in the StatCan list are the practical ones. They are reading and sorting text and sending and measuring marketing. They are also writing a workflow that did not exist and training the people who will run it. None of them is a promise that the tool will run the business. Reading text looks like an inbox, a pile of referrals or a supplier sheet. Sending marketing looks like a note to people who already bought from you. It is not a brand campaign with a film crew. A new workflow looks like a page the front desk can follow. The owner can be in a room with a patient at the time. Training looks like sitting with that person once on a real message. It does not look like a slide deck made for a big company.

The readiness note asks a different question. It asks whether this firm has the data, the process, the people and the decision in place. It does not ask about the average firm. Average firms are what surveys are for, but your firm is a counter, a desk and a schedule. If the readiness questions fail, the survey numbers will not save the project. If they pass, you still do not need the bigger number as a goal. What you need is the job.

A worked example

Take a clinic as the walkthrough. The owner saw the Chamber number and felt late, because friends in a group chat were already using it. The front desk still types appointment reminders by hand. The practitioner still scribbles the note after each visit, and billing still waits for someone to say the visit happened. Nothing in that room makes care through a model. The care is the practitioner. The model may come in at all. If it does, it comes in as a draft of a reminder. Or it comes in as a cleaner version of a note that the practitioner will still sign.

Walk the reminder first, because it happens every day and a wrong one is easy to spot. The front desk has the visit in the book. A draft can say the time, the place and what to bring. The front desk reads it, because some patients need a phone call and some need a short message. The draft does not know the difference. The practitioner is not in that choice unless the reminder holds a clinical instruction, and then the practitioner reads it. Billing does not need to see it at all. That is three roles and one job, and we named the handoff.

A shop down the street is even plainer, which is why the headline feels insulting there. The owner sorts supplier mail and customer questions in the same inbox. Sorting is text work, so a system can group the supplier sheets away from the sizing questions. The owner still decides which question gets an answer before lunch. The goods on the shelf were not made by a model. If the shop says it uses AI because the inbox got grouped, that is the Chamber kind of yes. If it said it uses AI to make the goods, that would be the StatCan question. The honest answer would be no.

A trades desk sits between the two. The lead tech’s site note is text, and turning it into a description the customer can read is a draft. The owner still sets the price, and the service is still the repair. You are near the production question only when the job truly depends on the system. A friend forwarding a chart does not count. Until then, hand over the draft and keep the price.

How we’d run it

First, write down which survey sentence you heard. If it was about generative use, say so in a full sentence. If it was about making the goods or the service, say that instead. Owners often mix the two into “most businesses are already doing this,” which is not what either source said. The groups are different, and the Canadian production share is the quieter fact.

Second, name one job on your floor that matches the practical list. It might be reading and sorting, or a message you send more than once. It might be a path that is not written down, or a person who has not been shown the path. Pick the one that already hurts this week. Do not pick the one that would look good in a post.

Third, write the roles the way we wrote them for the clinic reminder. Say who drafts, who must read, and who is left out. If everyone is involved, you do not have a handoff, because you have a meeting. Training means showing the person who will really run it.

Fourth, keep the output in a waiting place until the named person acts. A reminder waits, and so does a sorted inbox and a rewritten note. The send, the order and the published page stay with a person. They stay there until you can say what happens when they are wrong. For a clinic, wrong means a patient came on the wrong day. For a shop, wrong means a regular got a cold note. For a trades desk, wrong means a price the owner never set. If you cannot say what wrong looks like, you are not ready to let the draft travel.

Fifth, wait until that job has been dull for a while, and then ask the production question on purpose. Does the good or the service depend on this system? Or does a person still make the thing while the system handles the words around it? If a person still makes the thing, you are in the generative pattern. That is a fine place to stay. The smaller production share is not a scolding. It is a count of a harder change.

Sixth, check readiness for this firm and not for the average firm. You need data you trust and a process someone can say out loud. You need a person who will own the check, and a choice about what not to automate. The readiness note is that talk, and the pressure guide asks where the business really hurts. Neither one is a national percentage.

Common mistakes

The mistake we fix most often is blending the two surveys into one mood. The Chamber series shows a rise in owners who say they use generative AI. The Canadian series shows a much smaller share using AI to make goods or deliver services. It also shows what AI-using firms changed afterward. Those two are neighbors and not twins. If you quote one, you tell a different story than the speaker told. We have heard pitches that borrow the bigger number and then promise the smaller number’s result. That is how a draft tool gets sold as a new way to deliver the service.

The next mistake is treating a yes from an owner as a finished operation. An owner who tried a chatbot on a Sunday can say they use it. The front desk may never have seen the login. The survey cannot see that gap, because it asked the business and not each role. When we sit down, we ask the person who does the job and not the person who read the article. If those two answers differ, believe the doer.

Reading the tool shares as a quota is how small firms buy things they will not run. Text analytics and marketing automation are common among the Canadian firms that already use AI. But common is not the same as required. A clinic with no marketing program does not need a marketing project just so a share looks familiar. A shop that never posts a page does not need a content stack. The workflow share and the training share interest us more than the tool shares. They describe the work after the tool arrived. A firm that adopted a tool, never changed the path and never showed anyone is the firm we worry about. We worry even if its yes was confident.

Skipping the group is how 9,103 starts to sound like a club you failed to join. That count is responses from businesses that used AI. It sits inside a larger set of responses, in a survey with a set window of time. It is not a count of businesses that have figured out your counter. Use it to remember that the claim has a base, and do not use it as a mirror.

The last mistake is letting the headline pick the job. The owner wants whatever the article seemed to praise, but the floor wants the reminder or the inbox. We follow the floor. The article already gave us clean questions, and it does not get to staff the front desk.

What we’d do

We would tell the owner the bigger number and the smaller number in the same sitting. We would attach the question to each. We would also refuse to pick a tool until they can say which question they truly care about. Most of the time they care about a job and not about production. That is our opinion, and we think it is the kinder one. If you chase the production share before the reminder is reliable, you buy a system. The practitioner will then work around it.

We would hand over reading and sorting first, because the inbox is already there. We would keep a person on anything a customer could feel. We would write the workflow only for the job we just named. Then we would train the person who runs it before anyone else. A slide for the whole staff can wait, but the front desk cannot.

We would not use either survey as a score for the firm in front of us. Surveys measure everyone, and the readiness questions measure you. Write the job the way the customer feels it and name the person who checks it. Then ignore every percentage until that sentence exists. The percentages will still be true in the morning. Your counter is the thing that can slip.