A CIO analysis published October 2 makes an argument we see play out in practice: small businesses are having their AI moment, and it has exposed a services gap. Licences get bought. Structured daily use does not follow. The numbers it assembles are worth sitting with, because the gap between adoption and use is where most of the wasted money lives.
The numbers, with their dates
Bluevine's small business AI survey found 74 percent of owners using or actively testing AI, with 33 percent using it regularly across multiple areas, but 82 percent reporting barriers to deeper integration. Data security and privacy was the top barrier at 33 percent, up from 23 percent the year before. That survey covered 942 US small business owners and financial decision makers and was published in July 2026, so treat it as a few months old and US-based rather than a Canadian measurement.
Alongside it, a Business.com figure of 57 percent of US small businesses investing in AI, up from 36 percent in 2023, but only 30 percent of employees using it daily. And an analysis citing BlackFog research that 49 percent of workers use AI in ways their employer has not approved. Read those three together and the picture is clear enough: the tools are bought, the usage is patchy, and a lot of the actual usage is happening outside whatever was sanctioned.
For completeness on which tools, the same survey put ChatGPT at 57 percent of users, Gemini at 56, Copilot at 30 and Claude at 12 with steep year-over-year growth. We build this site with AI assistance and have no stake in which one you pick, so take that spread as market data rather than a recommendation.
Why licences do not become usage
Because a subscription is not a workflow. The common failure pattern looks like this: someone buys seats, announces them at a meeting, and expects adoption to follow from access. What people actually need is a specific answer to what do I use this for on a Tuesday. Nobody adopts a general-purpose tool in the abstract. They adopt a tool that shortens a task they already hate, and somebody has to identify that task, design the step, and show them.
The second failure is unmanaged risk pretending to be caution. When the business has no policy, the careful employees avoid the tool entirely and the rest use personal accounts. You end up paying for licences nobody touches while your client data moves through a free account you cannot see. That is the shadow AI problem we keep writing about, and it is the inverse of the usage number: adoption is not low, sanctioned adoption is low.
What the gap is actually made of
Three things, in rough order of impact. Workflow design: picking two or three recurring tasks and building the AI step into them properly, rather than hoping. Governance: a short written policy on approved tools and what never gets pasted in, which converts cautious non-use into confident use. Training: not a webinar, a working session where people do their own real task with the tool once, with someone watching. None of those are products, which is precisely why they are the part nobody sold you.
What to actually do
- Count your seats and your users. If you are paying for ten licences and three people use them weekly, you have a workflow problem, not a tool problem.
- Pick two tasks, not ten. Something recurring, time-consuming and low-risk if imperfect. Draft replies, meeting summaries, first-pass documents.
- Write the one-page policy so the careful half of your team is allowed to participate. Approved tools, what never goes in, who approves a new one.
- Run one working session per task, where people bring real work. Demonstrations of someone else's use case do not transfer.
- Check for the personal accounts. If half your team is using AI unofficially, the fastest adoption win is making the sanctioned path easier than the unofficial one.
- Measure something. Hours on a specific task, before and after. Without that, renewal is a guess and you will either over-buy or cut something that was working.
The takeaway
The honest version of the AI pitch for a small business has never been the model, it has been the decision about where it fits and the discipline to keep the risky uses out. That is advisory and implementation work, not another subscription, and the gap the research keeps finding is exactly that shaped. If you are paying for AI and not getting hours back, the thing to buy next is not more software.