Shadow IT used to mean the sales team’s Dropbox account or the spreadsheet that had quietly become the company database. It was a nuisance, but it was findable. Shadow AI is the same thing with two differences: it is much easier to sign up for, and it is much better at taking your data somewhere else.
A free chatbot needs an email address. A browser extension needs one click. An AI note taker can invite itself to a meeting because one attendee installed it. None of this shows up on a purchase order, so nobody in charge sees it happening.
What it looks like in a real office
When we look across a business of twenty to thirty people, the pattern is fairly consistent. A few personal ChatGPT or similar accounts, used for drafting and summarising. A browser extension or two that reads every web page the person visits so it can summarise them, which includes the pages inside your CRM and your accounts system. An AI meeting assistant that joins Teams calls and emails a transcript to everyone, including the client on the other end. Occasionally a “free trial” tool that was connected to the company Microsoft 365 account with a single click on a consent screen, and has been reading email ever since.
Nobody involved was doing anything wrong on purpose. They were trying to get through the day. That matters, because it changes what the fix should be.
Why this is a security problem, not just an IT tidiness problem
Three reasons, in order of how often they bite.
Your data leaves the systems you have secured. You have spent money on Microsoft 365 security, backup and, if you are certified, the Cyber Essentials controls. A personal AI account sits outside all of it. Whatever gets pasted in, from a client’s contract to a tender answer, is now held by a supplier you have no agreement with, under terms the member of staff clicked through without reading.
Consent screens are the new phishing. When someone connects an app to their Microsoft 365 account, they are granting it permission to read their email, files or calendar. Attackers know this and build convincing “AI assistant” apps whose only purpose is to be granted that access. Once granted, the app keeps it, even after the person changes their password.
Nobody can answer the question. Sooner or later a customer, an insurer or a tender panel asks how you use AI and what happens to their data. “We don’t know what our staff are using” is not an answer that wins work.
Why banning it does not work
The instinctive response is to block everything. We understand the urge. It does not work, for a simple reason: the tools are genuinely useful, and people will keep using them on their phones and their home laptops. All a blanket ban achieves is to push the usage somewhere you cannot see it and make people less likely to tell you when something goes wrong.
What works is the opposite. Approve a small number of tools, on business accounts, and make them easy to use. Then say clearly what must never go into any AI tool, and block the rest.
What good looks like
- An approved AI tools register. A short list of what staff can use, on which account, for what. New requests are assessed against a policy and approved, approved with conditions, or declined, rather than decided on the spot.
- App consent restricted to approved services. Microsoft 365 can be set so that staff cannot grant a third-party app access to company data on their own. Requests come to an administrator instead.
- Continuous discovery. Someone keeps looking for AI apps, browser extensions and cloud AI services in use, including ones people signed up to themselves. Unapproved ones are flagged, and blocked or restricted on managed devices.
- Data loss prevention. Alerts if sensitive information, such as customer records or financial data, is being shared with an AI service.
- A policy people have actually read. Briefed, acknowledged and tracked, with an annual review.
If you are working towards Cyber Essentials, or already hold it, this matters more. Unapproved AI tools sit outside your certified boundary. Bringing them inside, with approved tools managed like the rest of your systems, is the difference between a certificate that reflects reality and one that does not.
What this means for a business with 5 to 50 people
You are big enough for this to be a real exposure and small enough that nobody owns it. The good news is that at your size the whole thing is manageable: a policy, a register, a couple of Microsoft 365 settings and someone keeping watch. It is a few weeks of sensible work, not a transformation programme.
One thing to do this week
Ask your IT provider, or whoever administers your Microsoft 365, for a list of the third-party apps that have been granted access to your tenant. It is a standard report. If nobody can produce it, or the list is longer than you expected, you have found your starting point.
All of the above is included in our Protect & AI package: the policy your staff know about, the approved tools register, continuous shadow AI discovery with unapproved services blocked, restricted app consent, data loss prevention alerts and clear reporting on what was found, reviewed with you at your Clarity and Planning Reviews. Take a look at what is included and what it costs, so you know exactly what you would be getting.
