Blog

Copilot is secure, but is your data?

The rise of AI assistants is causing excitement in many organizations, but also a bit of chaos. After all, deploying AI is only effective if your data management is in order.

18 Nov 26
5
min read

Is Copilot actually secure?

Reinout Groothoff, Data Security & Compliance consultant, is frequently asked whether Copilot is truly safe: “First of all, Copilot is designed to be a secure tool. The risk lies in the data you feed into it.

Without strong data governance, you don't know what Copilot can see; with poor data security, it could, for example, inadvertently access documents or emails that were not intended for it.”

In practice, Reinout sees that organizations often want to enable Copilot without any policy: they want to hit the ground running. “But really, take the time to properly structure your data. Otherwise, you lose control over what information is shared with whom,” he urges.

Old and disorganized data

For instance, outdated or disorganized data regularly causes problems. “At one client, a document from 2005 was officially still considered current, leading to flawed decisions. And in shared folders, we found unclassified HR data, which allowed AI tools to accidentally process sensitive information.”

Shadow AI

Another example is ‘shadow AI’: employees using various AI tools outside official policy. That means without the knowledge or approval of the IT and security departments. “At one client, for instance, we discovered 63 different AI apps being used without any policy. Not with bad intentions, but simply due to a lack of clear guidelines from the organization.”

The biggest risk of shadow AI is that employees enter sensitive company data, customer information, or source code into public AI tools. This data can then be used to train the models, potentially giving competitors or third parties access to trade secrets and sensitive (personal) information.

These examples show how important it is to have your data management in order. This means cleaning up old files, shielding sensitive data, and making only relevant information available. That prevents AI from using incorrect or confidential data.

3 steps for a strong AI data policy: how to keep your data secure

Use this 3-step plan to first gain insight and structure in your data management, and only then deploy AI tools like Copilot.

Step 1: Clean up your data

The first step in a secure Copilot rollout is simple: know what you have. That means knowing where sensitive data lives, how it's being used, and who has access to it.

Reinout compares company data to an attic full of boxes: “You know there are valuable items in there, but you lack insight. You’re looking for something, but you have no idea which box it’s in or if you even still have it. Many organizations lack a clear overview of their data, making valuable information difficult to find.

He therefore advocates first creating visibility and structure in your data management before deploying AI tools like Copilot. “Think of those old photos you find in the attic,” Reinout adds. “Some you want to keep, others can go. But you want to decide for yourself what stays, and which photos you want to store with extra care. That is precisely what data classification does.”

With Microsoft Purview, you can set this up largely automatically. The tool identifies elements like contracts, social security numbers, or personal data and flags them as sensitive. “As a result, you reduce the risk of data breaches and more easily comply with regulations like the GDPR and the AI Act,” Reinout explains. “Purview displays a score to show where you stand. You literally see whether you are 60% or 90% compliant, and what steps are still needed.”

Data classification primarily provides visibility into what exists within your organization. You can also use Copilot without classification, but anyone who wants clear control over which data Copilot processes would do well to set this up first. “70 to 80% of data isn't confidential,” says Reinout, “but that 20% that is, you really need to protect properly.”

Such a classification process takes some time, but it delivers lasting clarity. “On average, it takes a few months, but you only have to do it right once.”

Step 2: Put locks on the right doors

Once it's clear where sensitive information resides, the next step is protection. This is handled with sensitivity labels, such as 'Confidential' or 'Internal only'.

“Labels are the locks on the doors of your house,” says Reinout. “You don’t want to put locks on everything, but you do want them on the room where your passports are kept.”

Labels can also prevent someone from copying or pasting sensitive information. “For example, if someone tries to copy a social security number from a document into a chat window, they'll receive a warning or the action will be blocked.”

In addition, you can link labels to departments so that HR only sees HR data and Finance only sees financial data. “This ensures that Copilot only gets access to information relevant to the task, and doesn't look where it doesn't belong.”

Step 3: Start securely and scale in phases

Many organizations don't want to wait months for AI. “They want to get started tomorrow,” Reinout smiles. “That’s understandable, as long as you approach it smartly.”

At OGD, we do it as follows:

  • Start with a small group: we test AI first in a small, controlled environment with key users, so everything can be tried out calmly.
  • Measure and learn: we use Purview reports and cloud discovery to see what data Copilot is using.
  • Phased scaling: only when everything is working well do we slowly expand usage throughout the organization.
  • Temporary solution: restricted content discovery: for quick starters, Copilot temporarily uses only documents that the user opens themselves, so that information remains within the correct context.

The human factor: adoption and awareness

A successful Copilot implementation depends entirely on the people using it.
“Technology is the easy part,” says Reinout. “The challenge is getting users on board. Ensure you have a solid adoption plan so employees understand why rules exist. Otherwise, they will bypass them and enter data into less secure AI tools. And then, of course, you’re back to square one.”

That is why we involve employees from various departments right from the start. This creates a widely supported policy that aligns with day-to-day practice. After the testing phase, we provide awareness training. Employees learn what labels mean and why they sometimes receive a notification.

Why this approach is worth it

Some organizations view security as a brake on innovation. But a well-structured data foundation ensures that Copilot works reliably and that employees can trust it. “AI is only as smart as the data you feed it,” says Reinout. “If your SharePoint says elephants are pink, then Copilot will say so too. So keep your data clean and up to date.”

A solid data policy also prevents confidential or sensitive information from accidentally being exposed. Think, for example, of medical records or internal processes. “You definitely don't want that on your record.”

AI requires a robust data policy. Working securely with Copilot is not a one-off project, but a continuous process. “You don’t have to—and can’t—arrange everything at once,” Reinout advises. “Start with insight, build your policy, and keep your data attic tidy in the meantime.”

This blog was previously published on ogd.nl

‍

Featured posts

We don't do off-the-shelf job openings. We create your ultimate role based on your personal ambitions and expertise.

Blog

When dependencies become sovereignty risks

Every CIO knows that an IT landscape has dependencies. These are usually managed through contracts, SLAs, and governance, which makes them feel manageable. The question is what happens when that assumption no longer holds true.

Blog

5 European AI alternatives for enterprise use in a changing AI landscape

In recent years, many organizations have built their AI capabilities on a small number of (primarily American) platforms. But the market has changed. We are increasingly seeing clients with explicit requirements regarding data sovereignty, risk diversification, contractual control, and the avoidance of vendor lock-in.

Blog

The impact of a cloud exit strategy on collaboration and communication

Recent geopolitical tensions and the growing focus on data sovereignty and privacy are causing increasing uncertainty for organizations. The question arises more and more frequently: what are the alternatives to Microsoft cloud services? And is it necessary, or even feasible, to develop a cloud exit strategy? This blog explores the realistic possibilities and limitations of such a strategy, focusing on one of the most critical business processes: collaboration and group communication.