Shadow AI at work is not a hypothetical. It is your bookkeeper pasting a customer list into a free AI tool on a Tuesday afternoon, because the list is messy and the tool cleans it up in nine seconds.
She is not being reckless. She is being efficient, which is what you hired her for.
However, that customer list has now left your business, and no one recorded where it went. Nobody approved the tool, nobody read its terms, and nobody can retrieve the data.
IBM published its 2026 Cost of a Data Breach report on July 29. One figure in it deserves fifteen minutes of your attention.
What shadow AI at work actually means
The term sounds technical. The reality is mundane.
Shadow AI is any artificial intelligence tool that somebody in your company uses for work without the business approving it. A free chatbot. A browser add-on that summarizes documents. A transcription service for meetings. An AI feature buried inside a tool you already pay for.
None of it appears on an invoice. None of it went through a review. So none of it shows up when somebody asks what systems handle your company data.
That is the entire problem in one sentence.
The number that changed this year
IBM found that unapproved AI tools featured in 43% of the security incidents studied. Last year that share sat near 21.5%, so it more than doubled in twelve months.
Two other findings sit alongside it. Close to seven in ten breached organizations had no policy governing AI use at all. And fewer than one in five had their governance people talking to their security people.
Now for the honesty this figure deserves. IBM studied 602 organizations between March 2025 and February 2026, and most of them are considerably larger than yours. The report’s headline cost, an average of $4.99 million per breach worldwide, is not what a 30-person company pays when something goes wrong. American organizations in the study averaged more than twice the global figure, which tells you the sample skews large and skews expensive.
So do not take the dollar figure home. Take the direction of travel instead. In one year, unapproved AI moved from an edge case to a factor in nearly half of incidents, and that shift is not confined to large companies. Small businesses adopted these tools faster, with fewer rules.
The tools are not the problem
Here is the finding that reframes the whole conversation.
Among organizations that suffered an AI-related security incident, 92% were missing basic access controls. No role-based permissions. No two-step login. The ordinary guardrails you would put on any system holding company data.
In other words, the incidents did not happen because artificial intelligence is dangerous. They happened because these tools arrived through the back door and nobody applied the normal rules to them.
That distinction matters commercially. Banning AI does not fix an access control problem. It just moves the same activity onto personal phones, where you have no visibility whatsoever.
What actually goes wrong
Picture three ordinary Tuesdays.
Your operations manager uploads a supplier contract to summarize it. That contract carries a confidentiality clause you signed. You have now shared it with a third party, which is precisely what the clause forbids.
Your sales lead pastes a pipeline export into a tool to draft follow-up emails. The export contains customer names, deal values and internal notes. Some of those customers made you sign a data protection agreement.
An employee leaves. Their personal AI account keeps every document they ever pasted into it, and you have no way to ask for it back. You do not even know the account exists.
The part owners underestimate
None of these three events is a hack. Nobody broke in.
That is why they escape notice for months, and why they surface at the worst possible moment: during a customer security questionnaire, an insurance renewal, or the due diligence phase of a deal you actually want.
At that point the question is simple and you cannot answer it. Which AI tools handle your company data, and under what terms?
Why banning it does not work
Every owner’s first instinct is a company-wide ban. It fails within a month, for a reason worth understanding.
The tools genuinely work. An employee who saves forty minutes on a report will keep saving forty minutes, whatever the policy says. A ban only guarantees that the activity stops being visible.
The businesses handling this well did the opposite. They named a short list of approved tools, applied the same access rules they use everywhere else, and told people plainly what may never be pasted into any of them.
That is a management decision, not a technology project. It costs you one afternoon.
The one-page policy that fixes most of it
Write a single page. Circulate it. Repeat it once a quarter.
Three approved tools, named. One rule about what never gets pasted anywhere: customer personal data, financial records, contracts, anything covered by a confidentiality clause, anything an employee would hesitate to email to a stranger. One named person to ask when somebody wants to use a new tool.
Nothing about this requires a technical background. It requires a decision, and the decision belongs to you rather than to your IT provider.
Questions to ask your IT provider
Ask your IT provider: can you tell me which AI tools currently have access to our company email, files or customer data?
Ask your IT provider: do the AI features inside the software we already pay for send our data outside the company, and can they be switched off?
Ask your IT provider: if an employee connects a new AI tool to our systems tomorrow, would anybody find out?
Ask your IT provider: do our approved tools use the same two-step login and permission rules as the rest of our systems?
Ask your IT provider: when a customer or an insurer asks which AI tools handle their data, what do we send them?
The last one is the test. If nobody can produce that answer within a day, you have a documentation gap that will cost you a contract before it ever costs you a breach.
What to take from this
Shadow AI at work grew from a curiosity to a factor in 43% of studied incidents in a single year. Meanwhile, 92% of the affected organizations had simply never applied their ordinary security rules to these new tools.
That is an unusually solvable problem. You do not need a platform, a project or a budget line. You need a named list, one rule about company data, and a person to ask.
Alchanis Technical Services is a veteran-owned security firm that works alongside small and mid-size businesses without a security team of their own. If you cannot name today which AI tools touch your customer data, that is the conversation to have. Visit alchanistech.com and we will build the one-page policy with you.
