Shadow AI is any use of AI that sits outside the systems and processes your business has approved. The picture that usually comes with the phrase is a member of staff choosing to paste a client contract into a free chatbot. That happens, and it is the version the security guidance is written about. In a small business, though, some of it never involved a choice at all. It is the AI feature that arrived inside a tool you already pay for. Nobody noticed because nobody signed up for anything.The job here is discovery. Find the AI already running inside the tools you approved, then make sure your policy actually covers it.
It is the AI version of shadow IT, where staff use software outside normal IT controls. The difference is that the software was already approved, and the AI arrived later.That makes it harder for a policy alone to catch. Knowing where these features appear matters more than relying on a detection tool.Microsoft does offer a tool designed to find them, but it requires an E5 licence.
Where shadow AI hides in a normal office
Walk through a working day and the places are easy to spot once you know the shape.
The meeting that transcribed itself. A colleague connected a meeting assistant to their calendar months ago. It now joins every call they are invited to, including the one with your biggest client, and sends a summary to an account nobody else can see. Teams, Zoom and Google Meet each have recap features too. The Teams version needs a Copilot or Teams Premium licence, while the Zoom and Meet versions depend on admin settings that nobody may have opened.
The writing assistant in the browser. Grammar and rewriting extensions can read the text entered into the pages where they run. That may include your CRM, HR system or client portal.The level of access depends on the permissions granted when the extension was installed. If a vendor later adds generative rewriting, that feature can operate under the same permissions the original spelling checker already had.
The assistant inside the chat app. The messaging app a sales team uses with customers now has an AI assistant built into the conversation list. It is a tap away in a thread that contains order details and phone numbers.
The AI hiding inside everyday work
The design and document tools. Presentation builders, PDF readers and image editors have each added generation and summarisation features. A PDF reader that can summarise a contract may be sending the contract to the vendor’s service to do it; the product’s own documentation says whether it does, and whether it keeps a copy.
The free assistant already in the suite. Microsoft 365 includes Copilot Chat for business users without a separate Copilot licence. Google also includes Gemini within Workspace plans.These options usually offer better governance than consumer AI tools. There are also practical ways to use them without creating unnecessary risk.The challenge is perception. Staff often see them as ‘the company AI’, so anything pasted into them can feel automatically approved.
The desktop app on one laptop. ChatGPT, Claude and similar tools all have installable desktop versions. One person installs one, it reads their screen or files when asked, and on a laptop that is not managed by the business it may leave no trace anywhere an administrator looks.
None of these people did anything wrong. That is the point. Much of what turns up in a small firm is a sequence of sensible-looking upgrades and settings, which is why an inventory finds more than a policy does.
Why it matters even when the tool is reputable
The NCSC’s September piece on the hidden risks of shadow AI focuses on larger organisations, but the risks apply more widely.
When data flows into consumer AI services, you can lose visibility and control over it. That can increase the risk of a data breach, lost intellectual property and regulatory non-compliance. In practice, for a small firm, that last one means personal data ending up somewhere your privacy notice never mentioned. The NCSC’s line that you cannot manage what you do not know is the whole problem in one sentence.
For a business with clients, the sharper risk is confidentiality. A meeting summary sitting in a personal account, or a draft contract summarised by a consumer tool, is a client’s information held somewhere the client never agreed to. Whether the vendor is reputable is beside the point if nobody in your firm knows the data went there.
The NCSC is careful to say that it does not recommend anyone stop using AI. Neither is The Unite Group. Staff reach for these tools because they help, and a firm that bans them gets the same behaviour on personal phones with less visibility. The job is to see what is in use and bring it inside the approved list.
What a Business Premium tenant can and cannot see
Microsoft has added a Shadow AI page to the Microsoft 365 admin centre.It is still in preview, and the prerequisites will rule out many small businesses.To view it, you need a Microsoft 365 E5 licence. Devices must also run Defender for Endpoint, while managed Windows machines need Intune enrolment.As of its August update, the tool can detect seven desktop AI agents. These include the ChatGPT and Claude desktop apps. It can also block one of them. It says nothing about browser extensions, meeting bots or AI features inside the apps you already use.
Business Premium still has a visibility gap
Entra ID P1 includes a cloud app discovery feature, which supports the conditional access capabilities in Business Premium. However, the feature relies on firewall or proxy logs that you need to upload first. Small businesses rarely have this configured.In practice, that means most Microsoft 365 Business Premium firms still lack an automated view of the AI tools staff are using. Waiting for a perfect discovery tool is unlikely to help. What it does have is useful.
- The app consent list in Entra. Third-party tools that staff have connected through their Microsoft 365 sign-in, meeting assistants and calendar bots among them, appear under enterprise applications in the Entra admin centre, with the permissions each was granted. Reading that list does not take long and is worth doing this week.
- User consent settings. Two settings in the same place, turning off user consent and switching on the admin consent workflow, mean an administrator has to approve any new app before it can read mail, calendar or files. That stops the next meeting bot arriving unannounced.
- The question. Ask each team what they use and why, with no consequences attached. The NCSC’s advice is to build a culture where staff talk openly about the tools they need, and in a firm of thirty people a straight question is the quickest route to a complete list.
What to do this quarter
- Start with the consent list and the conversation, then write down what you find.
- Decide which tools are approved, which are fine with conditions, and which have to go.
- Put that list in front of staff alongside the AI policy points that need to exist before anyone uses these tools. Where the approved option is the Copilot already in Microsoft 365, say so, and staff have something to move to.
Then keep the register alive.A quarterly review of AI features is a simple way to catch new defaults before they become a problem. Add the review to the licence and security checks already on the agenda. The Unite Group’s managed IT services include quarterly reviews, making them a natural place to cover this. It is also the simplest control on this page.