A 2026 survey by PagerDuty and Wakefield Research found that sixty six percent of office professionals had used AI tools at work while believing those tools were not permitted. More than a third of them admitted entering customer data into public models while doing it. Separately, a 2026 survey by the security firm Anagram found fifty eight percent of employees had pasted sensitive material, including client records and internal documents, into large language models.
Read those two findings together and the picture is not one of reckless staff. It is one of people who found something that made them faster, and who were never told where the line was.
The gap is worst at the smallest businesses
This is the part that matters for a business in its first year. A 2026 Founder Reports survey of more than two thousand American workers found that forty four percent said their employer had no clear AI policy, or that they were not sure whether one existed. At companies with fewer than ten employees, that figure rose to fifty nine percent.
Meanwhile adoption keeps climbing. The US Chamber of Commerce puts generative AI use among small businesses at fifty eight percent, up from forty percent the year before. Thryv found adoption jumping to sixty eight percent among firms with ten to a hundred employees. Microsoft reported that seventy five percent of workers use AI at work and that seventy eight percent of those bring their own tools, frequently without approval.
The two curves are moving in opposite directions. Use is rising fast. Governance at the smallest businesses is not moving at all.
Why this is a first year problem specifically
A large company discovers shadow AI through an audit. A three person business discovers it when a client asks why their contract terms appear to be known to somebody who should not know them, or when a piece of work goes out with a confidently invented figure in it.
There is a second reason it bites harder at this size. Everything at a first year business is learned by observation. There is no handbook, no training week, no compliance function. Whatever you do, the next person copies exactly, and they copy it within their first fortnight. If you paste a client email into a consumer account to get a quick summary, that becomes the standard, and nobody will ever mention it out loud.
The silence is the actual danger. People are not hiding AI use because they enjoy breaking rules. They are hiding it because asking feels like admitting they need help, and because no one has told them it is fine.
Two settings and one sentence
The remedy here is unusually cheap, which is what makes the neglect so odd.
Start with the accounts. Consumer tiers and business tiers of the same product frequently have different data retention terms, and the difference is whether your input can be used to train the underlying model. Business and team tiers generally exclude it by default. Personal accounts frequently do not. Checking that setting on every account in use takes about ten minutes and is the single highest return action available.
The credential risk compounds this. Security researchers found more than two hundred and twenty five thousand ChatGPT related credentials for sale on dark web markets during 2025, most harvested from compromised personal devices. An employee using their own account on their own laptop is a route into your client information that no business subscription would have created.
Then write down what never goes in. Not a policy document. A list. Client names, addresses, financial details, contract terms, anything told to you in confidence. Six lines is enough, and the specificity is what makes it usable. A rule saying to be careful with confidential information is not a rule, because everybody already believed they were being careful.
The sentence that does most of the work
Tell people that using these tools is expected and that nobody needs to hide it.
That single statement changes the behaviour the surveys are measuring. The employees pasting client records into public models are not doing it in defiance of a policy. They are doing it in the absence of one, on personal accounts, precisely because the alternative is a conversation they would rather not start.
Permission plus a short list of exclusions produces better outcomes than prohibition, because prohibition at this scale is unenforceable and everybody involved knows it.
Training is where the return actually sits
There is a figure worth sitting with. Deloitte found that small businesses which trained employees on AI tools saw two point three times higher productivity than those which deployed the same tools without training, on an investment of four to eight hours per person.
That is a striking ratio for something most businesses skip entirely. The tools are cheap and the training is free, and the gap between a business that spent an afternoon on this and one that did not is larger than the gap between a business using AI and one that is not.
It also explains a finding that otherwise looks strange. A recent index reported that ninety six percent of creative organisations have AI policies and that ninety six percent of staff ignore them. Policies without conversation are documents. The conversation is the mechanism, and the document is a reference for afterward.
What to verify, every time
The other half of this is quality, and it is simpler than the data protection half.
These tools produce confident text that is sometimes wrong. Not obviously wrong, which would be manageable, but plausibly wrong in ways that survive a quick read. Facts, figures, names, dates, prices, and anything legal or tax related all need checking against a source before they leave the business.
The useful pattern is that AI is strong where you can judge the output and weak where you cannot. Drafting something you will rewrite is a good use. Producing a figure you have no way of verifying is not. If you would not notice that the answer was wrong, that is precisely the task to be careful with.
Where this leaves a first year business
Three things, none of which take longer than an afternoon.
Set up business accounts and check the training setting on each. Write down the six things that never go into any tool. Tell whoever works with you that use is expected, that hiding it is the only actual problem, and that everything gets checked before it leaves.
Do it before the first hire rather than after. Rewriting a culture is considerably harder than defining one, and the person who joins next month will copy whatever they find on their first day, including the parts you would not have chosen.
The surveys will keep reporting that most small businesses have no policy. That statistic is not describing a failure of technology or budget. It is describing an afternoon that nobody got round to.