Your AI Provider Doesn't Train on Your Data. So What's the Risk?

Your AI provider has confirmed that your business data won't be used to train its models. That's good news. But does it mean your organisation's information is fully protected?

Not quite.

As generative AI becomes part of everyday working life, IT leaders are rightly asking more questions about how information is handled.

Where does the data go? Who can access it? What happens to the information employees enter into prompts? And can it be used to train future AI models?

These are all important questions.

But there's a risk in focusing so heavily on whether AI providers train on customer data that we overlook what happens to that information elsewhere.

Because even when the answer to the training question is reassuring, there are other things worth looking at.

First, let's clear up the confusion around AI and data

One of the biggest misunderstandings around generative AI is that all AI services handle information in the same way.

They don't.

And the distinction isn't simply whether someone is paying for the service.

Take ChatGPT, for example.

With consumer plans, including paid individual subscriptions such as Plus and Pro, conversations may be used to improve models depending on the user's data settings.

Business and Enterprise workspaces operate differently. OpenAI states that business customer inputs and outputs aren't used to train its models by default.

Microsoft Copilot also provides enterprise data protection for eligible organisational use, with prompts, responses and Microsoft Graph data excluded from foundation-model training.

These are meaningful protections, and they matter when organisations are deciding which AI services to approve.

But three things are worth separating:

• Encryption helps protect information while it's stored or transmitted.
• Model-training commitments determine whether customer information can be used to improve underlying AI models.
• Access controls and governance determine who can use information, which systems can access it and what happens when it's shared.

They're connected, but they're not interchangeable.

A provider can offer strong encryption and a clear commitment not to train on customer data, while an organisation still has risks within its own environment.

The information AI can access matters just as much as what happens to it

Let's take Microsoft Copilot.

One of its strengths is that it can work with information already held across Microsoft 365, helping employees find, summarise and make use of content they have permission to access.

But there's an important detail here.

Copilot respects existing permissions. It doesn't automatically fix permissions that were too broad in the first place.

Imagine a SharePoint site that started life as a small project workspace.

Over time, more people were given access. Documents were added. Some contained commercially sensitive information. Nobody had a particular reason to revisit the permissions because everything seemed to be working.

Now introduce AI that can help employees discover and summarise information much more efficiently.

The underlying access problem hasn't changed.

What's changed is how easily someone with existing permission might find information they didn't even realise was available to them.

That's not an argument against Copilot. It's an argument for understanding your data estate before introducing new ways to interact with it.

If your permissions, sharing arrangements and information ownership haven't been reviewed for a while, AI gives you another good reason to do it.

Then there's the AI your IT team doesn't know about

Most IT teams have a reasonable understanding of the technology they've approved.

It's the technology they haven't approved that can be harder to account for.

An employee discovers an AI tool that produces excellent presentations. Someone else uses a browser extension to summarise meetings. Another team starts experimenting with an AI-powered service that promises to save hours of administrative work.

Individually, these decisions may seem harmless.

People are trying to be productive, after all.

But each new tool raises questions.

What information is being uploaded? How is it processed and retained? Is it shared with other services? What access has the employee granted? And what contractual protections apply?

This is where shadow AI becomes a practical governance issue.

Not because employees are necessarily doing anything wrong deliberately, but because AI adoption can move much faster than the processes organisations use to assess and approve new technology.

The answer isn't simply to tell everyone to stop experimenting.

It's to understand what they're using, why they're using it and whether the business can provide an appropriate, approved alternative.

What about the information people put into prompts?

Even with an approved enterprise AI service, employees still need to understand what information is appropriate to share.

A request to summarise a publicly available report is very different from uploading a spreadsheet containing customer records, confidential pricing or employee information.

And although enterprise services can offer stronger contractual and technical protections, that doesn't mean every type of information should automatically be entered into every available AI feature.

The organisation still needs to consider the purpose of the processing, the information involved, applicable data protection obligations and the controls available within the particular service.

Practical guidance makes a difference here.

Employees shouldn't have to interpret a lengthy policy every time they want to use AI.

They need to know which tools are approved, what types of information they can use with them and when they should stop and ask.

We'll explore that human side of AI governance in our follow-up article, How Do You Get Employees to Actually Follow Your AI Policy?

Don't forget the connections behind the scenes

An AI service rarely operates in complete isolation.

It might connect to document repositories, customer systems, email, collaboration platforms or third-party applications.

These integrations can make AI significantly more useful.

They can also introduce additional considerations around permissions, data flows and accountability.

For example, an AI tool might have permission to retrieve information from a connected system, while an automated workflow could potentially take action based on that information.

Those are different levels of access and risk.

It's worth understanding what each integration can read, what it can change and whether a person needs to approve an action before it happens.

The same applies when existing software suppliers introduce new AI features.

Just because an organisation has approved a supplier doesn't necessarily mean it has assessed every new AI capability that supplier subsequently adds.

So, what should IT leaders be looking at?

The starting point isn't necessarily another security product or another lengthy policy.

It's getting a clearer picture of how AI interacts with the information, systems and people already within your organisation.

A few useful questions to ask:

1. Do we know which AI services employees are actually using? Include both approved platforms and services adopted independently.
2. Have we reviewed the information our approved AI tools can access? Pay particular attention to overshared files, permissions and sensitive data.
3. Do we understand our providers' data-handling commitments? Look beyond model training to retention, processing, integrations and administrative controls.
4. Are employees clear about what they can share? Guidance should be specific enough to apply during a normal working day.
5. Are appropriate technical safeguards in place? Consider access restrictions, data protection controls, monitoring, audit capabilities and incident response.
6. Who is responsible for keeping this under review? AI services and their capabilities are changing, so governance can't be a one-off exercise.

The aim isn't to make AI difficult to use.

It's to give people the confidence to use it productively, without creating risks the organisation hasn't properly considered.

The real question isn't just what your AI provider does with your data

A commitment not to train on your business information is important.

It's something organisations should understand, verify and value when assessing AI services.

But it's only one part of a much bigger picture.

The risks worth paying attention to often sit closer to home.

The permissions that haven't been reviewed. The tools nobody knew were being used. The information shared without enough thought. The integrations that have quietly gained access to more than they need.

And the policies that look perfectly sensible until someone tries to apply them on a busy Tuesday afternoon.

AI can bring real benefits to a business. The challenge is making sure the way it's introduced, accessed and governed keeps pace with those opportunities.

At Robinscroft, we believe that starts with asking the right questions, understanding what's already in place and taking practical steps to make technology work securely for the people using it.