
Shadow AI: what your staff already use, and how to see it
Shadow AI is the AI your staff already use without anyone signing it off: a chatbot in a browser tab, an extension that summarises every page, a note-taker that joins meetings. Writing a policy achieves little while you do not know what is running. This article is about that first step: making visible what is already there, without turning it into an interrogation.
Shadow AI needs no installation: a browser tab or an extension is enough. That is why it never shows up in your software list.
You make it visible from five sources: network traffic, cloud app discovery, app consents, browser extensions and expense claims. And by simply asking.
Sort what you find by what data goes in, not by brand. What is allowed is your organisation’s decision.
Shadow AI: what it is and why it is already running
Shadow AI is a form of shadow IT, with one difference that makes it harder: nothing needs installing. Someone opens a chatbot in the browser, pastes in a quote or a customer email and has a better text within a minute. Nothing was bought, and nothing appears in the software inventory.
It happens for the same reason as all shadow IT: the work goes faster, and the approved alternative either does not exist or is more awkward. That is not disobedience; it tells you what people need.
So the risk is not the tool but what goes into it. Customer personal data, a contract, source code or an HR file in a personal account with a provider you have no agreement with: that is the real problem. The ICO’s guidance on generative AI makes the same point about personal data.
Where you can see which AI is being used
No single place shows everything, but five sources together give a good picture. You probably have most of them already:
- Network traffic. Logs from your firewall, DNS filter or web filter show which AI services are visited from the office and from managed laptops, and how often.
- Cloud app discovery. Security suites such as Microsoft Defender for Cloud Apps recognise traffic to known cloud apps, AI services included, and give each one a risk rating.
- App consents. Microsoft Entra ID or the Google Workspace admin console lists which apps staff have granted access to their mail, calendar or files. That is where the AI note-takers and writing assistants connected with one click turn up.
- Browser extensions. On managed devices you can list installed extensions. AI extensions that read every page stand out at once.
- Expense claims and card statements. A subscription someone pays for and claims back is an AI tool in business use with no agreement behind it.
Note what you do not see: use on a personal phone or home laptop. That is why the next step matters as much as the technology.
Asking without it becoming an interrogation
The quickest way to find out what staff use is to ask them. It only works if the answer has no consequences for what has happened so far. Say so up front and mean it.
- Ask about the purpose, not just the tool. “What do you use it for?” tells you which need an approved alternative has to cover.
- Ask what data goes in. Text without names, customer data, internal documents, code. That determines the risk more than the name of the service.
- Ask whether it is a personal or a business account. For many services that changes what the provider may do with the input.
- Keep it short and anonymous where you can. Five questions get more honest answers than twenty.
Put the answers next to what the technology showed. Where they differ, the gap is usually use outside the office.
What to do with the list: sort by data, not by brand
A list of AI tools only becomes useful once you sort it. Do that by what happens to the data, because the same service can be fine with a business account and not with a personal one.
- Allow. A service you have a business agreement with, where input is not used to train the model and you know where the data is held.
- Allow with conditions. Business account only, no personal data, or only after a short briefing.
- Replace. A service with no agreement that sensitive data does go into. Block it only once the alternative is ready, or use simply moves to personal phones.
Four questions per service are enough: is input used for training, where is data stored, who at the provider can access it, and is there a data processing agreement if personal data is involved.
The hardest kind: AI that reads and listens along
A chatbot someone pastes text into is visible and limited: only what goes in can come out. Harder are the tools that have been given standing access.
- AI note-takers that add themselves to every meeting in the calendar, including meetings with customers who know nothing about it.
- Writing and email assistants with read access to the whole mailbox.
- Browser extensions that read the content of every open page, internal systems included.
That access stays in place long after the employee has forgotten the tool, and sometimes after they have left. Revoke consents nobody can explain, and set things up so staff can only connect apps from verified publishers, or only after admin approval. It is one setting with a lot of effect.
Keeping it visible instead of one clean-up
An inventory is out of date within months; new AI services appear every week, and new AI features arrive in software you already use. What helps is a fixed rhythm:
- A monthly look at new apps in the traffic and new app consents.
- An approved alternative that is just as easy. As long as the official route takes more effort, the other one wins.
- A place to request a new tool, with a quick answer. A request left for three weeks is an invitation to sort it out yourself.
- Guidance by role. Since February 2025 the EU AI Act requires organisations to ensure sufficient AI literacy among those who work with AI; UK organisations serving EU customers meet the same question. What belongs in a policy is in our AI policy template article.
Where our work stops
What we do. We make visible which AI services appear in traffic, app consents and extensions, restrict who can connect apps, revoke unexplained access and set up an approved alternative, such as a business AI environment inside your own tenant.
What we do not do. We do not read what staff type into a chatbot, and we do not assess individual employees. We are not lawyers: whether a particular use is lawful under data protection law or the AI Act, and what rules apply to monitoring staff, is for your lawyer or privacy adviser. The ICO publishes guidance on monitoring workers; settle that question before we switch anything on.
Questions we get about this
The five that come up most often.
What is shadow AI?
Shadow AI is the use of AI tools by staff without the organisation approving it or knowing about it: a chatbot in the browser, an AI extension or a note-taker connected to the calendar. The risk lies mainly in the data that goes in, such as customer data or internal documents, with a provider you have no agreement with.
What is the difference between shadow AI and shadow IT?
Shadow AI is a form of shadow IT, but usually nothing needs installing. A browser tab is enough, so it does not appear in your software list. And it is less about a device or program than about data ending up with a third party.
How do you find out which AI tools staff are using?
Combine firewall or web filter logs, cloud app discovery in your security suite, the list of app consents in Entra ID or Google Workspace, browser extensions on managed devices and expense claims. Then ask staff directly, with no consequences for what has already happened.
Should we ban ChatGPT and other chatbots?
A ban without an alternative usually moves use to personal phones, where you no longer see it. Better is an approved service with business terms, clear rules on which data may go in, and blocking what is left only once the alternative is in place.
Do you read what staff type into a chatbot?
No. We show which services are used and which apps have access, not what individual employees enter. The rules on monitoring staff are for your lawyer or privacy adviser to assess.
Further reading
Where shadow AI connects.
Want to know which AI is already in use at your organisation?
We put traffic, app consents and extensions side by side and show which AI services are running and what data may be going there. What is allowed is then your call.
Practical IT knowledge in your inbox
New guides on management, security and the workplace, written by the people doing the work. No sales talk, and you can unsubscribe in one click.