Someone on your team probably used an AI tool for work this week, on an account you do not control, without mentioning it. Not out of malice - because it helped them hit a deadline.
The numbers say this is now the norm, not the exception. Verizon's 2026 Data Breach Investigations Report, drawing on 858,440 data-loss-prevention events involving generative AI tools, found that regular AI use tripled in twelve months from 15% to 45% of employees. Two-thirds of that activity ran through personal accounts the employer could not see.
This has a name - shadow AI - and most of what's written about it is aimed at corporations with security teams. Small businesses have the same problem with none of the visibility, so let's deal with it at the right scale.
What actually leaks, and why it matters
The data flowing into unapproved tools is not exotic. Per the DBIR's dataset: source code, client proposals, HR records, financial documents. In a small business that translates to: the pricing spreadsheet pasted in "to make a summary", the client contract uploaded "to check a clause", the salary list "to draft the review letters".
Three things make this worse than ordinary sloppiness. First, you can't see it - personal accounts leave no trace in your systems, so your exposure is unknown even in principle. Second, the data may persist - depending on the tool and account type, what's pasted in can be retained and used outside your control; free personal tiers generally offer the weakest guarantees, and those are exactly what shadow use runs on. Third, your obligations don't pause - UK GDPR duties, recently reshaped by the DUAA, and client confidentiality agreements apply just as fully to data an employee pasted into a chatbot as to data on your server. "We didn't know" is not a defence you want to test; the UK survey data shows how thin current practice is, with only 24% of AI-adopting businesses having any security processes around AI at all.
One more finding worth knowing: more than 15% of users in the DBIR dataset had unauthorised AI browser extensions installed - tools that can read the pages your staff browse, including internal systems. Extensions deserve a line in your policy of their own - not least because malware now routinely arrives dressed as a well-known AI app or extension, and where a tool gets installed from matters as much as which tool it is.
Why the ban reflex fails
The instinctive response - "no AI tools at work" - fails for a simple reason: the tools work. Your team uses them because they genuinely save an hour here and a day there. A ban doesn't remove the incentive; it removes the visibility. Work shifts to personal phones and home laptops, and your data goes with it - you've traded a manageable risk for an invisible one, and you've told your most productive people that initiative is a compliance problem.
The realistic goal is different: make the sanctioned route more convenient than the shadow one.
The one-afternoon policy
This genuinely fits on one page, and writing it is an afternoon's work. Five sections:
1. Approved tools, business accounts. Name the small set of tools your business sanctions - for most small firms, one general assistant and perhaps one or two role-specific tools is plenty. Paid business tiers, not free personal ones: the account terms are the product here - business tiers typically offer data-handling commitments the free tiers don't. The cost of a few seats is trivial next to what it replaces.
2. Red-line data. List what never goes into any external AI tool regardless of account: client-identifiable information, staff personal data, financials, credentials and keys, anything under NDA. Plain words, five bullet points, no jargon. The test we suggest: if you wouldn't email it to an outside consultant without a contract in place, don't paste it.
3. Work happens on work accounts. The single highest-value rule, because it converts invisible risk into visible use. Pair it with the flip side: the business pays for the tools, so nobody has a reason to use their personal account.
4. A fast yes/no channel. Shadow AI thrives where asking permission is slower than just doing it. Name one person to whom "can I use this new tool?" goes, with a 48-hour answer. Most answers will be yes - say so in the policy, it changes the tone from prohibition to process.
5. Blame-free disclosure. Someone has already pasted something they shouldn't have. You want to hear about it, which means the policy must say - explicitly - that reporting it is thanked, not punished. Then you can actually respond: rotate the exposed credential, note the exposure, adjust.
Sign it, share it, revisit it quarterly - the tool landscape moves fast enough that an annual review is a dead document.
Treat adoption as a signal
Most shadow-AI coverage frames employees as the risk. We'd frame them as the signal: 45% of your workforce adopting a technology without being asked is the clearest product-market fit imaginable. The businesses that will get hurt are not the ones using AI - they're the ones officially not using it while actually using it invisibly. A one-page policy is not bureaucracy; it's the cheapest way to keep the upside and see what you're exposed to. And candidly: helping businesses take exactly this step from invisible to deliberate AI use is part of what we do - but the policy above needs no vendor. Write it this week.
Sources
- Verizon - 2026 Data Breach Investigations Report (2026)
- Verizon - Data Breach Investigations Report resource centre (2026)
- GOV.UK - Cyber Security Breaches Survey 2025/2026
- ICO - UK GDPR guidance
Want a second pair of eyes on your AI policy - or help picking the approved-tools shortlist for your team? That's where our AI implementation work usually starts. We'll help you set it up properly.
