The AI Access Problem Most European Businesses Have Not Checked For

In 2023, just 8.1% of European Union businesses with at least ten employees were using artificial intelligence (AI), according to Eurostat. Two years later, that figure had reached 20%. The change has been even faster in some corners of the economy. By 2025, 55% of large EU businesses were using AI, up from 30.5% just two years earlier.

For entrepreneurs and business leaders, the speed of that adoption creates a problem that is easy to miss. It is one thing to know that people inside your company are using AI. It is another to know what those AI tools can see once they are connected to the rest of the business.

Ofer Klein, CEO and co-founder of Reco, an agentic security company, says companies often discover AI operating in more places than they expected.

“A business may have approved a small number of AI tools, but then they discover AI features inside meeting apps, browsers, collaboration platforms, CRM systems, support tools, and workflow applications,” Klein told Entrepreneur Europe.

The bigger surprise, he says, is often what those tools can access. Some are connected through ordinary user permissions or OAuth grants, meaning an AI application may be able to reach information that the employee who connected it can access.

For European founders who are racing to put AI to work, that turns a productivity decision into a security, privacy, and increasingly commercial question.

The AI you approved may not be the AI you have

Shadow IT is hardly a new problem. Employees have long downloaded software or signed up for cloud services without going through their company’s formal procurement process. AI, however, makes that problem even more difficult to see.

A July 2026 investigation published by the International Association of Privacy Professionals reported that DataGrail had examined data-protection assessments from 2,400 popular business software providers advertising AI capabilities. It found that 63.6% did not properly disclose all AI subprocessing activity. Nearly a third of the AI systems examined disclosed that they were involved in sensitive-data processing or automated decision-making.

There is another layer to the problem. AI does not always arrive as a separate application with ‘AI’ written across the front. It can come as a new feature inside software employees already use.

Klein says Reco has seen business teams enable an AI assistant for something as narrow as summarizing meetings or following up with customers, then discover that the application could also reach shared folders, CRM notes, contracts, or support records because of the permissions already sitting behind it.

“The team thought it had approved a productivity feature,” he says. “In practice, it had created a new access path into customer or employee data.”

That distinction becomes particularly important in Europe because much of the information sitting behind those applications may be personal data.

GDPR still applies

The arrival of agentic AI does not create a loophole in Europe’s existing privacy rules. In February, Spain’s data protection authority became one of the first European regulators to publish detailed guidance addressing agentic AI and data protection. Among the issues it highlighted were inadequate access management and insufficient human oversight.

The UK’s Information Commissioner’s Office has reached a similarly practical conclusion. Its guidance says companies need a clear reason for the personal information an agent can access. Giving an agent broad access because the information might prove useful later is not enough. The regulator points instead toward carefully choosing the databases and tools an agent can reach, alongside measures such as system permissions, monitoring, and human approval where appropriate.

That’s close to the problem Klein says companies are finding in practice.

An employee may have a legitimate reason to access a customer record, contract, or personnel file. It does not automatically follow that every AI application connected to that employee should have the same reach. And increasingly autonomous agents complicate the question because they can use the information they have access to while completing tasks across different systems.

The Confederation of European Data Protection Organisations warned in a July paper that agentic systems capable of pursuing complex goals with limited human intervention raise questions about GDPR, accountability, transparency, and human oversight.

For entrepreneurs and business leaders, the useful question is therefore less abstract than “Are we GDPR compliant?” It is: What can every AI system inside this company actually access?

Customers now want to know

There is another reason businesses should be able to answer that question. Their customers increasingly want to know, too. Klein says enterprise buyers are asking vendors more detailed questions about AI use, data access, retention, security controls, and how models interact with company information.

“Procurement teams want evidence, not just reassurance,” he says.

For a startup technology vendor that is trying to sell into a bank, insurer, or large European enterprise, that can move AI governance from the compliance department directly into the sales process.

Martin Woodward, global legal director and global responsible AI officer at Randstad, described another difficulty in comments recently published by IAPP. Traditional IT relationships are often relatively straightforward. With shadow AI, one piece of software can involve numerous underlying technologies and providers.

That makes discovering everything in the chain harder. A company can even find itself acting as a provider or deployer under the EU AI Act without having fully understood the AI use case it has brought into the business, Woodward warned.

The timing matters. Transparency obligations under Article 50 of the EU AI Act began applying on August 2, 2026. They apply to certain AI systems rather than every company that happens to use AI, and the obligations differ depending on whether a business is providing or deploying the relevant system.

GDPR and the EU AI Act, therefore, should not be treated as interchangeable rules. But both make it increasingly difficult for a business to take the position that it simply does not know where AI is being used.

Start with the boring part

The temptation is to solve this with another AI policy. That is useful, but it does not tell an entrepreneur or business leader what is actually happening after someone clicks ‘connect.’

Klein argues that companies should begin with a live inventory of AI agents, tools, copilots, agents, and AI-enabled features. Then they need to map the accounts, permissions, data sources, and workflows connected to each one.

From there, someone in the organization has to determine what kind of data is involved, whether the level of access makes sense, and who inside the company owns the tool. That inventory also has to change when employees, permissions, and software change.

Vincent Rezzouk-Hammachi, a partner at international law firm Bird & Bird, made a similar point when discussing shadow AI with IAPP. Companies, he said, are struggling with who should actually own AI governance because it cuts across so many parts of the organization.

That may be one reason the problem survives even in companies that already have an AI policy.

Klein says the most common mistake he sees is treating governance as an approval exercise. A company approves an application, writes down the rules, and considers the job finished. But what an AI tool can do may change after approval, as employees connect it to more files, accounts, and workflows.

For an entrepreneur or business leader, there is a fairly simple test. Pick an AI tool your company uses and ask four questions. What can it access? Why does it need that access? Who is responsible for it? And what happens when you decide it should no longer have access?

If those questions are difficult to answer, the problem is already more concrete than another policy can fix.