Only 9% of Canadian Small Businesses See AI Gains

The 2026 Zensurance Small Business Confidence Index, released September 14 and based on a national survey of 1,000 Canadian small business owners, reports that 32% say AI has had little to no impact on their business and another 27% are not sure how it even applies to what they do. Just 9% credit AI with meaningful productivity gains or cost savings. Seven per cent say it has hurt their revenue or client base. We sell AI consulting, and those numbers describe our own market accurately.

The numbers

One more finding deserves attention. Asked about plans, more owners intend to introduce new products or services (26%), expand into new markets (22%), or invest in new tools and equipment (20%) than to learn about or invest in AI (15%). That is a rational allocation from people who have not seen a return, rather than a failure of imagination.

Read the caveats before quoting the number

This is one survey of 1,000 respondents, self-reported, published by an insurance company as part of its own brand research programme and distributed by press release. None of that makes it wrong, and the direction matches other adoption research. It does mean no single percentage in it should carry a strategy.

Self-report is the specific limitation to hold onto. A firm whose bookkeeper quietly saves three hours a week may not attribute that to AI. A firm that bought seats nobody opened may still count as having adopted it. The survey measures what owners believe about their businesses, which is worth knowing and is not the same as measuring what happened.

The press release also reports the one-in-four positive figure as being among those who felt an impact, while the 9% and 7% figures are stated without the base being spelled out. Those may or may not be drawn from the same subset, so the arithmetic between them should be treated loosely.

Why no impact is often the correct answer

A two-person landscaping company that bought a chatbot subscription in 2025 and never used it has honestly experienced no impact. So has a machine shop whose work is physical, whose quoting is a phone call, and whose paperwork is four invoices a week.

Some businesses have little AI-shaped work in them right now. Telling those owners they are behind is a sales tactic rather than an observation, and the 15% who would rather spend on equipment or a new market are frequently making the better call for their business this year.

The category worth separating is the 27% who are unsure. That group usually does have automatable work and cannot see it, which is a different problem with a cheap fix.

What the 9% did differently

The survey does not say, so this is our observation from the Canadian businesses we work with rather than a finding from the data. Four things recur.

They picked one repeated task, not AI in general. Writing quotes. Summarising site visits. Pulling numbers off supplier invoices. Adoption framed as a task has an owner and a finish line. Adoption framed as a technology has neither.

They timed it before and after. Actual minutes across a handful of instances, recorded by the person doing the work. Without a baseline, the answer to whether a tool helped is a feeling, and feelings about new software are unreliable in both directions.

They changed the process, not just the tool. Drafting a quote in two minutes saves nothing if it still waits three days for a signature. The businesses that got a return moved the bottleneck, which usually meant a decision rather than a purchase. The method is in process mapping before you automate anything.

They kept a person on the output. Verification on anything reaching a customer or the books. That sounds like it cancels the saving and does not, because checking a figure takes a fraction of the time that finding and typing it does.

The 10% who do not trust the tools are right

Ten per cent said AI tools are not reliable or accurate enough to trust. That is a well-founded position rather than a misconception, and it is more fixable than it sounds.

Accuracy turns out to be largely a property of how the task is framed. In the ExtractBench study we covered earlier this month, the same model scored 85.4% on a 13-field extraction and 0% on a 369-field one. Ask for less in one go, verify per field, and the reliability problem becomes a design problem, which we set out in extracting data from PDFs with AI.

If you are in the 32%

Pick the task in your week you most dislike doing and that you do most often. Time yourself on it five times. Spend two hours trying to do it with whatever assistant you already have access to. Time five more.

If the difference is under about 15%, you have a real answer and can stop reading about AI for six months with a clear conscience. If it is larger, you have found your first candidate, and the next steps are documenting it properly, covered in standard operating procedures an AI can follow, and deciding what the freed hours are for, covered in capturing the hours AI saves.

Either outcome beats another year of being unsure. The broader Canadian picture is in the AI adoption gap in Canada, and the ROI calculator turns your timings into an annual number.

Frequently Asked Questions