In October, I went to a local AI event – I’m currently using Claude to improve various in-house processes at Innovate, so I was interested in what others are up to. While I was suitably impressed by what I saw, one thing became abundantly clear when I chatted to some of the presenters: While AI certainly opens the doors for those with non-technical backgrounds to develop apps etc., it’s the lack of understanding how the code works that opens a Pandora’s box of cybersecurity risks – and most of my conversation partners weren’t aware of said risks.
Granted, the use cases at the event were already advanced, but it shouldn’t detract from the fundamentals. Whether you use Copilot, Claude, ChatGPT or any other AI tool – for many, it’s become part of their daily working lives. What prompted me to write about data protection is because of mainly one aspect: visibility.
Why Is AI a Data Protection Risk?
You may have introduced AI in the form of Copilot licences (or any other tool of your choice). But simultaneously – unless you’ve blocked it – your staff may be using a private instance of an AI assistant on a free plan. So – do you even know where your data goes?
If I may ask you one question: Can you reliably tell that no-one in your organisation touched an AI tool with your company data before the introduction of an acceptable use policy?
Or let me go even further back: do you have an acceptable use policy in the first place (you can download one here)?
Why it’s so important: you and your staff should know what can and what shouldn’t be pasted into an AI tool – approved or not. But the general risk is that once your data is pasted into an unapproved/unconfigured tool (like a free plan e.g. ChatGPT) – the data leaves your control, and you’re effectively handing it to a third party, so everything that happens to it is governed by their terms.
This is what’s known as Shadow AI. And I’m not surprised that it’s happening, given that AI use is encouraged seemingly everywhere – why wouldn’t they use it to make their lives easier. Unfortunately, many people simply don’t realise that pasting something into a chat window means sharing it with a third party.
But – you don’t even have a contract or data processing agreement in place, which is a clear compliance gap. You’re also lacking an audit trail. Regulatory exposure (UK GDPR) is a real risk not only in terms of fines from the Information Commissioner’s Office but also reputational damage and loss of client trust.
Your goal obviously shouldn’t be to stop AI use entirely, but to give your team the tools to do so safely, with clear rules, approved tools, and the right technical controls.
A Shadow AI Case Study: Confidential Patient Data Uploaded
In September 2026, it transpired that a staff member at Johns Hopkins Bayview Medical Center in Baltimore had uploaded a spreadsheet with information about more than three thousand patients to a public, non-approved AI tool. An absolute no-go. While there’s no reported misuse of the data, it still involved the user deleting the chat, and the hospital contacting the vendor to delete the data and reaching out to all affected patients – not a good look.
And all of that was a result of one well-intentioned upload. Deleting the chat afterwards also doesn’t guarantee the data is gone; at that point, you’re relying on the vendor. The hospital reportedly already had guidance telling staff not to put clinical records into unapproved AI tools, which shows that a policy on its own isn’t enough. Now – while you may not be dealing with patient records, I’m sure you wouldn’t want that happening to your client list, HR files or financials.
You can download a Shadow AI checklist here.
It Starts With What Goes Into the Chat Window
Before we get to the technology, your first line of defence is your people. A simple traffic-light system works well because staff can apply it in seconds:
- Green: public or non-sensitive information – fine for approved AI tools.
- Amber: internal or confidential information – approved enterprise tools only.
- Red: personal data, client-confidential information, credentials or source code – never, unless a tool has been specifically approved for it.
Encourage a quick ‘pause before you paste’: Is this an approved tool? Does it identify a person? Could I anonymise it or use dummy data instead? And remember, it’s not just typed text – file uploads, screenshots, browser extensions and meeting transcription bots all count.
Finally, make it easy to own up. If someone pastes the wrong thing, you want to hear about it as soon as possible – so create a no-blame route for reporting mistakes.
Make the Safe Route the Easy Route
As the economy has shown time and time again: banning something doesn’t make it go away. Instead, appoint an owner – ideally your IT director or a technical lead – with a clear mandate and backing from the top. Publish a short acceptable use policy that explains what’s allowed, what isn’t and why. But also provide an approved tool that’s not only easy to use but that you also provide training for.
Before approving anything, check the basics: is there a data processing agreement, where is the data stored, how long are prompts retained and are they used to train the model? Where personal data is involved, a data protection impact assessment (DPIA) is a sensible step under UK GDPR. Yes, enterprise licences cost money – but far less than a reportable breach.
For Your IT Lead: Access, DLP and Visibility
A data policy is only worth the paper it’s written on unless it also has practical (in this case technical) implications. If you’re the one taking the lead on implementation, here’s where I’d focus.
Find out what’s already in use. Review proxy, DNS and firewall logs for traffic to AI services, and use the app discovery in your CASB or secure web gateway (Microsoft Defender for Cloud Apps, for example). Check OAuth consents and browser extensions too. A short anonymous (do make sure it’s actually anonymous, though) staff survey will tell you which tools people find useful – and that should shape your approved list.
Lock down user access. Use enterprise tenants only, with single sign-on through your identity provider, MFA and conditional access so AI tools are only reachable from managed devices. Where possible, block personal account sign-ins to AI services on company devices. And if you’re rolling out Copilot or a similar assistant that inherits user permissions, fix oversharing in SharePoint, OneDrive and Teams first. This is admittedly something you should be doing on a regular basis anyway, as teams and roles – and hence the data required to access – change over time. Reason being is that AI will surface whatever a user can already access but may currently not even be aware of.
Put data loss prevention in place. Start with sensitivity labels so your data is classified, then use endpoint and browser DLP (Microsoft Purview, for example) to detect and block sensitive data being pasted or uploaded to unapproved AI sites. Build in a graduated response – audit, then warn, then block – and run in audit mode first so you can tune out false positives before enforcing.
Configure and monitor. In approved tools, switch off training on company data, keep retention to a minimum and restrict plugins and connectors. Send audit logs to your SIEM, alert on DLP events involving AI destinations and have an incident playbook ready. Under UK GDPR, a reportable breach must be notified to the ICO within 72 hours – so you need to know quickly when something has gone wrong.
Where to Start
Data protection in the age of AI is a people and process challenge first, and a technology challenge second. To get started, do the following as soon as possible: name an owner, find out which AI tools are currently in use, and publish a one-page policy (you can download one here). Then back it up with the right access controls and DLP over the following months.
Not sure where you stand? Contact us to arrange a review of Microsoft 365 security setup.