Is it safe to give company documents to AI?

This is one of the questions we hear most often. The worry is understandable: contracts, prices, customer data. The short answer is that it can be safe, but it depends on three things: what account you use, what the AI can see and what the contract says. Here is how to check them.

Personal or business account

The biggest risk is usually not the AI itself but an employee pasting a company document into a free personal account. In consumer versions conversations may be used to train the models. In ChatGPT this is on by default and can be switched off in the settings. In Claude's personal plans, since 2025, the user chooses whether to allow it. Either way, the decision belongs to the employee, not the company.

Business versions work the other way round. OpenAI states that it does not train models on data from its business plans and API by default. Anthropic does not train models on content from its paid business services. Microsoft states that prompts, responses and data in Microsoft 365 Copilot are not used to train the foundation models.

The takeaway is simple: give people a business tool and a clear rule that company data never goes into personal accounts. A ban without an alternative usually just hides the use.

The AI sees what you see

When AI is connected to your documents it usually respects the existing access rights. Microsoft, for example, says Copilot only surfaces data the user has at least view permission for. That is good, but it holds a surprise: if the payroll folder is shared with the whole company out of old habit, the AI will find it in seconds.

So review permissions before connecting anything. It is the cheapest security step in the whole project.

Where the data goes

If AI processes personal data, the provider is usually a processor under the GDPR. You need a contract under Article 28, which the large providers offer as a data processing addendum (DPA). Also check where the data is physically processed. Microsoft, for example, processes prompts from EU customers within the EU, but lists exceptions such as Anthropic models in Copilot.

Transfers to the US currently rely on the EU-US Data Privacy Framework. The EU General Court upheld it in 2025, but an appeal to the Court of Justice is still pending. If your data is especially sensitive, processing in the EU or on your own servers reduces this risk.

Risks specific to AI

Hidden instructions. If AI reads external documents or emails, they may contain hidden text that tries to give it commands. OWASP ranks this attack first among the risks for applications built on language models. Sensitive information disclosure is second. The answer is that the AI should have no more rights than the task needs, and risky actions should go through a person.

Wrong answers. AI can sound confident and be wrong. Important answers should point to their source so a person can check.

A short checklist

Business accounts for everyone using AI with company data. A signed data processing agreement. A clear rule about what never goes into AI, such as passwords and sensitive personal data. Access rights reviewed before connecting. A record of who asked what and what they got. For high-risk processing of personal data, an impact assessment under Article 35 of the GDPR.

If you use AI only for public information and drafts of marketing copy, the risk is small and these steps may be overkill. This article is a general overview, not legal advice. If you would like us to check your setup together, write to us.