The 2026 US Chamber of Commerce survey landed a headline number that has been doing the rounds ever since: 89% of small businesses now use AI in some capacity, up from 36% in 2023.
Taken at face value, that is one of the fastest technology adoption curves in history. Faster than the web, faster than smartphones, faster than email. Nine in ten small businesses, in three years.
Then you look at a different set of numbers and the story falls apart. The US Census Bureau’s Business Trends and Outlook Survey, which asks businesses whether they actually use AI to produce goods and services, puts genuine production use at 17 to 20%. JP Morgan Chase Institute went further and studied real transaction data rather than survey answers. Their figure: 17.7% of small businesses showing payments to AI services.
So which is it? Are 89% of small businesses using AI, or 18%?
Both, and the gap between the two numbers is the most useful thing in either survey.
A definitional gap, not a contradiction
The surveys are not contradicting each other. They are measuring different things, and the difference is exactly the difference that matters for your business.
Pasting an email into ChatGPT to tidy up the tone counts as “using AI” in the Chamber’s survey. So does asking a chatbot to summarise a document, or generating a social media caption once a fortnight. It is real usage, and it is not nothing. But it is dabbling. It lives in a browser tab, it depends on whoever remembered to do it, and if it stopped tomorrow the business would not notice.
The Census question is harsher. Is AI part of how you produce your goods and services? Is it wired into the actual operation of the business? By that standard, roughly one business in five qualifies, and the JP Morgan transaction data, which measures money actually leaving accounts, lands in almost exactly the same place.
The 70-point gap between “we use AI” and “AI is part of how we operate” is where nearly every small business currently sits. They have adopted the tools without adopting the change.
The UK picture rhymes with this. UK SME adoption sits around 35 to 39%, up from 25% in 2024. Lower headline numbers than the US, but the same shape underneath: steady growth in businesses touching AI, with a much smaller core using it in anger.
Why dabbling feels like progress
Dabbling is seductive because it produces visible moments of value. The email got written faster. The proposal got summarised. Everyone in the Monday meeting agrees that AI is impressive.
But none of those moments compound. The time saved on one email does not change your cost base. The summarised proposal does not shorten your sales cycle. Individual, ad-hoc usage produces individual, ad-hoc gains, and they evaporate the moment the individual gets busy or leaves.
A business that “uses AI” in a browser tab has adopted a habit. A business with AI wired into its workflows has adopted a capability.
This is why so many owners privately report the same experience: we use AI all the time and honestly, nothing much has changed. The revenue line looks the same. The headcount pressure is the same. The bottleneck that was killing them in 2024 is still the bottleneck. They are in the 89% and it feels like being nowhere.
The businesses that crossed the gap
Here is why the gap is worth taking seriously rather than shrugging at: the businesses on the other side of it are reporting materially different outcomes.
Salesforce found that 91% of AI-using small firms report revenue increases. The US Chamber’s own research found AI-adopting small businesses are 2.3 times more likely to report revenue growth than non-adopters.
Now, treat those numbers with appropriate care. Businesses that invest in embedding AI are often the better-run businesses to begin with, and self-reported revenue gains deserve a raised eyebrow. But the direction of the evidence is consistent: the reported gains cluster around businesses that moved AI into their actual operations, not businesses that occasionally prompt a chatbot.
What does “embedded” actually look like at SME scale? Concrete, unglamorous things. Enquiries that get classified, drafted and routed automatically instead of sitting in an inbox. Quotes generated from your actual pricing rules rather than rebuilt from scratch each time. Customer records that update themselves from emails and calls. Content pipelines that run on your data and your tone rather than a blank prompt box. AI connected to the systems where the work already happens.
None of that comes from buying a better tool. Which brings us to the uncomfortable part.
It is a systems problem, not a tools problem
The instinct, when AI is not delivering, is to go shopping. A better model, a new subscription, the tool a competitor mentioned. But the businesses stuck at the dabbling stage almost never have a tools problem. The tools are astonishing and mostly interchangeable.
What they have is a systems problem. Their data is scattered across inboxes, spreadsheets and disconnected apps. Their processes exist in people’s heads rather than in any form a machine can plug into. There is no one accountable for wiring anything together, so every AI interaction starts from zero, in a browser tab, and ends there too.
You cannot embed AI into workflows you have never defined, or connect it to data you cannot get at. That is the real barrier, and it explains why the gap has persisted for three years even as the tools got dramatically better. Better models do not fix disconnected systems.
What to do about it
If you suspect your business is in the 89% but not the 18%, here is the practical route across.
Audit what “using AI” currently means in your business. List every place AI is actually touched, by whom, and for what. If everything on the list is manual, individual and optional, you are dabbling. That is a starting point, not a failure.
Pick one workflow, not one tool. Choose a single process that is repetitive, rule-based and currently eats hours: enquiry handling, quoting, reporting, content. Depth in one workflow beats shallow usage across ten.
Fix the plumbing first. Before any AI work, ask whether the data that workflow needs is accessible. If it lives in one person’s inbox or a spreadsheet on someone’s desktop, that is the first job. Embedded AI is only ever as good as the systems underneath it.
Wire it in so it runs without heroics. The test of embedded AI is that it works when the enthusiast is on holiday. If a process depends on someone remembering to open a chatbot, it is not a system yet.
Measure a business number, not an activity number. Not “prompts per week” but hours recovered, response time, quotes issued, revenue per head. The Census would count you when AI changes how you produce, so measure production.
Then, and only then, expand. One embedded workflow teaches you more about what AI can do for your business than a year of dabbling. Repeat the pattern on the next bottleneck.
The 89% number tells you the tools have won the argument. The 18% number tells you the work has barely started. That 70-point gap is not a scandal, it is an opening, and it will not stay open forever.
Flux Dynamics helps SMEs move from dabbling with AI to embedding it, wiring it into the workflows, systems and data where the gains actually live. Start a project with us if you are ready to get past the browser tab.