Is Your Company's Data Safe When You Use AI?

«We're launching internal projects to work with artificial intelligence»: it's a phrase we hear often, with a mix of enthusiasm and, at times, scepticism. But it only looks at the official initiatives, the ones with a budget and clear objectives. AI data security in the enterprise has already been at stake for a while, because artificial intelligence came in through the back door: in employees' browsers, in conversations with ChatGPT or similar services. Free, at least on the surface.

The price, almost always, is data. In the first weeks of a project, we often ask the companies we work with a simple question: can you tell us what your employees pasted into an AI tool this month? Most of the time the answer is an embarrassed silence, followed by a «but who cares about my data anyway?». No one steals that data: the company, without realising it, has already handed it over itself.

This article directly answers, without alarmism, a question almost every company asks itself and almost none addresses openly: is your company's data safe when you use AI?

The most common case: an employee discovers that ChatGPT or Copilot saves them an hour a day, and starts using it to summarise emails or draft contracts. No one has forbidden it, so they do it. Meanwhile IT hasn't activated anything, management isn't aware of it, and the company policy, when it exists, dates back a few years.

This is the phenomenon known as shadow AI: everything artificial intelligence does in a company without the company knowing about it. The term echoes the shadow IT of ten years ago, with one substantial difference: an AI tool doesn't just run on an unauthorised server. It reads and processes whatever it's fed, and in some cases it retains it.

The Italian numbers show just how unbalanced the phenomenon is relative to the ability to govern it: in 2025, 71% of large Italian enterprises had launched at least one structured AI project, but only 9% had formal governance for the use of these tools. 84% had purchased licences for generative AI solutions, but just 24% had explicitly banned the use of non-company tools, and only 19% of employees said they used exclusively approved tools.³

Here company size matters little; timing matters. Adoption arrived before the rules. When that happens, what AI can or can't do with company data is decided by the tool's vendor, by default, or in the worst case by the individual employee, prompt after prompt, without management knowing.

Training. Some AI tools, especially on free or consumer plans, may use conversations to improve future models, unless the user explicitly configures otherwise. A snippet of code, a piece of company or client data could, in theory, influence the answers that same tool will give other users in the future.

Retention. Even without training, data still remains stored for a period ranging from a few days to about two months, depending on the provider, for security purposes, on servers the company doesn't directly control and whose geographic location is often not clearly disclosed.

Operational use. This is the least risky part, but not therefore harmless: once data leaves the company, the company loses control of it. It can no longer track it according to its own policies, nor delete it on request.

A concrete example: in spring 2023, some Samsung engineers pasted portions of proprietary source code into ChatGPT to help with debugging. A work habit, not a cyberattack — one that today millions of employees replicate every day.⁴

The phenomenon is accelerating: the share of sensitive company data shared with AI tools went from 10.7% to 34.8% in two years, with a sharp jump visible in the last year alone, from 27.4% to 34.8%.⁵

The phenomenon has stopped being an isolated behaviour or the carelessness of a few: it has become the main channel through which confidential information leaves companies — more than email, more than the USB drives that were once the nightmare of IT managers.

On 30 March 2023 the Italian Data Protection Authority (Garante) ordered a temporary restriction on OpenAI's processing of Italian users' data, citing the lack of a legal basis for training the algorithms and of age-verification systems. It was the first temporary block of a generative AI in Europe.⁶

A few weeks later, the service resumed thanks to measures introduced by OpenAI, but the investigation continued, ending with a corrective order and a €15 million fine for failing to promptly notify a data breach and for unlawful processing.

This is a decisive issue for Italian SMEs: those who use consumer plans like ChatGPT Plus or Claude Pro for individual employees often don't have a valid corporate DPA, because the terms of service exclude it. The risk is twofold: breach of the GDPR and of the provider's own terms.⁷

In the same period, AI-driven attacks grew by 56%, and breaches traced back to a model-inversion attack — the technique that exposes the data used to train a model — cost on average $6 million: a confirmation of how much more complex protecting data becomes once artificial intelligence enters the picture.

These numbers hit close to home: from 2 August 2026 the European AI Act entered its enforcement phase, and for the most serious violations — including non-compliant use of high-risk systems — it provides for administrative fines of up to €35 million or 7% of global annual turnover, whichever is higher. Reduced thresholds apply to SMEs and startups, but the underlying sanctioning logic remains unchanged.⁹

The AI Act also introduces an obligation in force since 2 February 2025 for every company, regardless of size: to adopt measures to promote employees' AI literacy, so that they know how to recognise which data must never be entered into an AI system.¹⁰ Since July 2026 the obligation has been made more proportionate to the company's size and risk level, but it remains valid for everyone, SMEs included.¹¹

A consumer plan, whether ChatGPT Plus or another equivalent personal service, is designed for individual use. An enterprise plan, on the other hand, comes with extra elements, designed specifically for corporate use and its policies:

It matters little to ask whether AI is safe in the abstract. What matters is asking whether this specific system is, with these settings, for this use, and with this data.

Treating cybersecurity and privacy as secondary details, leaving the configuration of corporate plans and staff training until after the AI tools have already been activated, is a strategic mistake. Those who postpone these decisions think they're saving time and resources, but they're actually choosing to discover and face the problem at the worst possible moments, when it will be too late.

There's no need to give up AI: you just need to never take it for granted, whatever tool it's built into.

The latest Italian data show widespread delay on this front: 76% of Italian SMEs haven't invested, and don't plan to invest, in artificial intelligence in the near future, and only 7% have launched structured training programmes for their staff. In most cases the problem is a lack of clarity over who, within the company, decides on AI-related issues, including data security — more than caution.¹³¹⁴

Using AI at work is the easy part. The part that takes real work is always knowing where information ends up once it leaves: a provider with a verified data processing agreement, and staff trained on what should never be shared.