Is AI safe? How to keep control over your own data

In brief

AI is not safe or unsafe in itself. It depends on the agreements: where your data is stored, who can access it, whether it is used for training and how long it is kept.
A free consumer service usually gives you none of those agreements. A business service under a data processing agreement does.
There are three common routes: a public service, a business cloud service with agreements, or an open model on your own servers.
Put the answers in the contract, not just in a brochure.
At Prudai, documents and data are stored in Hetzner's EU data centres: databases on our own servers, document files in Hetzner Object Storage. Chat requests and the processing of knowledge-base documents run through EU endpoints of Google and Microsoft.

Is AI safe for confidential data?

Yes, as long as you stay in control of your data. A language model (a large language model, or LLM) processes whatever you give it: a question, a document, a case file. So the question is not whether the model is "safe". The question is what happens to that input.

Four questions determine the risk:

Where is the data? In which country and with which party is it stored and processed?
Who can access it? Can the supplier's staff or other customers see it?
Is it used for training? Does the supplier use your input to improve models?
How long is it kept? Are questions and answers stored, and for how long?

If you do not know those answers, you do not know whether you comply with the GDPR either. An organisation that uses AI on personal data remains the controller. The supplier is a processor (Article 28 GDPR), and that requires a data processing agreement.

Where is the risk with public AI services?

With a free consumer service, the provider's terms apply, not yours. You have no data processing agreement, you often do not know in which country the data is stored, and whether your input is used for training depends on settings that each user has to check.

That is fine for a recipe or a holiday plan. It is not fine for a client file, a medical report or an HR file. In healthcare, law firms and government, "the provider promises" is not enough. You have to be able to show it.

Which routes lead to safe AI?

There are three common routes. They differ in control, cost and management effort.

Route	Who runs the model?	Agreements about your data	Suitable for
Public service (free or individual subscription)	The provider	The provider's terms, usually no data processing agreement	General questions without confidential data
Business cloud service through an EU endpoint	The cloud provider	Data processing agreement, choice of region, agreements on training and retention	Most organisations with confidential data
Open model on your own servers	Your organisation or your managed service partner	Everything in your own hands	Organisations with their own IT team and very strict requirements

Business cloud service with agreements

The major cloud providers offer language models in a business environment. Microsoft states that for its models in Azure, prompts and completions are not available to OpenAI and are not used to train models. With a standard deployment they are processed in the geography the customer chooses. Google states that its AI platform does not use customer data to train models without permission, and explains which settings are needed to keep nothing stored (zero data retention).

Read the details. With a "Global" deployment, Microsoft may process data in any region where the model runs. For some features, such as grounding with Google Search, Google keeps logs for a short period. Your settings therefore decide the outcome.

Open model on your own servers

You can also run an open-weights model such as Llama, Mistral or Qwen yourself, on your own hardware or in a closed cloud. You then decide everything yourself: access, logging and updates.

That control comes at a price. You need serious computing power (GPUs), people who manage and secure the model, and you often lag behind the best models. For most organisations this does not outweigh a well-configured business cloud service.

Which agreements should you make with your AI supplier?

A brochure is not a guarantee. Put these points in the data processing agreement or its annex:

Checklist of five agreements about your data with an AI supplier: where the data is stored, no training on your data, how long questions and answers are kept, who has access, and which sub-processors are involved.

Location: in which countries is the data stored and processed?
Training: the supplier does not use your data to train models.
Retention: how long are questions and answers kept, and can that be zero?
Access: who at the supplier can see the data, and under which conditions?
Sub-processors: which parties are in the chain, and will you be told when that changes?

Also ask about certification of the management system, such as ISO/IEC 27001, and for healthcare NEN 7510. The strategic side of this choice is covered in Private AI in Europe.

How it works at Prudai

Prudai does not run language models on its own servers. We weighed the routes above and chose the business cloud service, with storage in our own hands:

Storage: customer documents and data are stored in Hetzner's EU data centres, in Germany and Finland: databases on Prudai's own servers, document files in Hetzner Object Storage. No hyperscaler sits in between.
Language models: chat requests and the processing of documents in knowledge bases run through EU endpoints of Google Vertex AI and Microsoft Azure OpenAI (resource in Sweden Central). If a model fails in the chat, another model takes over, also through an EU endpoint.
Training: it has been agreed that the model providers do not use your data to train their models. Prudai does not use your data to train models either.
Retention: the model providers do not keep prompts and outputs permanently. To speed up responses they keep a short-lived cache (for Gemini in working memory only) that expires within minutes and at most 24 hours. If abuse is suspected, a provider may temporarily retain prompts and have them reviewed by authorised staff. That is why we do not promise zero data retention.
Certification: Prudai is certified to ISO/IEC 27001:2022 and NEN 7510-1:2024. The certificates cover the management system of Prudai B.V.

The full list of sub-processors and the security measures are in our Trust Center.

For the specialist

Processor and sub-processor. An AI supplier that processes personal data on your behalf is a processor. It may only engage other processors with your prior written authorisation (Article 28(2) GDPR). A general authorisation is allowed, but then it must inform you of changes.
Transfers outside the EEA. If data leaves the European Economic Area, the transfer rules of Chapter V GDPR apply (Article 44 onwards). An EU endpoint reduces that risk, but also ask about administrative and support access from other countries.
DPIA. If you use AI for sensitive personal data or at scale, a data protection impact assessment is often mandatory (Article 35 GDPR).
Settings, not just contracts. At Microsoft, the processing location depends on the deployment type (standard, DataZone or Global). At Google, zero data retention depends on settings such as request-response logging, the use of grounding and an exception to abuse monitoring.

Frequently asked questions

Is ChatGPT safe for business data?

That depends on the version and the agreements. The free consumer version gives you no data processing agreement and limited control. Business versions and cloud services do offer contractual terms on training, retention and location. Check them before you enter confidential data.

What is private AI?

Private AI is a set-up in which your organisation decides where data is stored, who can access it and whether it is used for training. That can be done with your own model on your own servers, but also with a business cloud service under strict agreements.

Does my data stay in the Netherlands?

Not automatically. Ask your supplier in which countries the data is stored and processed, and put it in writing. At Prudai, documents and data are stored in EU data centres in Germany and Finland, and chat requests and the processing of knowledge-base documents run through EU endpoints.

Is my data used to train AI?

With the business services of Microsoft and Google, not without your permission, according to their own documentation. At Prudai, your data is not used to train Prudai or AI models. Check with every supplier that this is in the contract.

Sources

What generative AI is and how to check whether it is right is explained in What is generative AI?.

Would you like to know how Prudai protects your data? Visit our Trust Center or see Prudai's AI colleagues.

Updated on 6 October 2026

Is AI safe? And how do you keep control over your own data?