Is my data used for AI training? How to check providers

Whether your data is used for AI training depends on the provider, on the version you choose and on the contract – there is no blanket answer. As a rule of thumb: consumer and free versions often reserve the right to use data for training, business versions usually exclude it by contract. For your company, only what has been agreed in writing counts. This article shows the five points you should check with every AI provider before company data goes into the tool.

In brief

• Whether input is used for AI training is decided not by the technology but by the contract – and that differs considerably between free, personal and business versions.
• A training exclusion is the contractual commitment that your documents and input will not be used to improve models.
• An opt-out in the settings is weaker than a contract: it can change and is hard to prove.
• Data that has once gone into a model can practically no longer be retrieved – which is why it pays to check before the first upload.
• The five checkpoints: data processing agreement (DPA), explicit training exclusion, the right version, settings, sub-processors and storage period.

Why is AI training with company data a problem in the first place?

If a provider uses your input to train its models, you give up some of your control over that data. This affects two levels at the same time.

Data protection: company documents almost always contain personal data – customer names, employee data, contact details. The GDPR requires purpose limitation: data you collected to perform a contract may not simply be used for entirely different purposes such as training someone else’s AI models. If the provider pursues its own purposes with the training, it also leaves the role of a processor bound by your instructions – and this is exactly what the data processing agreement (DPA) must exclude or clearly regulate.

Trade secrets: cost calculations, recipes, customer lists and internal know-how are often more valuable than any machine in the workshop. What has gone into a model can practically no longer be removed in a targeted way. Even if a model as a rule does not reproduce input word for word: you have given up control over what happens to this knowledge.

For context: this article sets out general principles and does not replace legal advice in an individual case.

Consumer or business: what is the difference?

The same provider can behave completely differently in its free version than in its business version. The typical differences:

Feature	Consumer/free version	Business version
Use for training	Often reserved, sometimes with an opt-out in the settings	Usually excluded by contract
Data processing agreement (DPA)	Usually not provided	Offered, often as part of the contract
Account management	Private individual accounts, no control by the company	Central management, rights and rules set by the company
Storage location and deletion	Hardly any influence	Can be regulated by contract
Proof when something happens	Difficult	Contract and documentation available

The table also explains the biggest everyday risk: employees who work with private free accounts because the company does not provide anything official. Then the consumer terms apply – without management knowing about it.

The five checkpoints in detail

1. Is there a data processing agreement (DPA)? As soon as the provider processes personal data on your behalf, a data processing agreement under Art. 28 GDPR is mandatory. If a service does not offer one, it is unsuitable for company data. Templates and guidance are available from the Austrian Economic Chamber, WKO (wko.at).
2. Is the training exclusion explicitly in the contract? Look for a clear wording that input and documents will not be used for training or to improve models. Vague statements (‘we take data protection seriously’) do not count.
3. Are you really using the right version? Check which plan is actually in use – and whether all employees work through the company accounts. The best business contract is of no use if staff use private free accounts on the side.
4. How are the settings configured? Business environments also have switches for telemetry, history storage or product improvement. Check them once centrally, document them and set them sensibly.
5. Who are the sub-processors, and how long is data stored? Behind an AI service there are often further service providers, such as model operators and data centres. Check their list as well as how long input remains stored and when it is deleted.

Which questions should you put to the provider in writing?

If the documents remain unclear, a short written enquiry helps. Four questions have proven useful:

• Are our input and uploaded documents used to train or improve AI models – even in anonymised form?
• Where is the data stored, and who can access it from where?
• Which sub-processors are involved, and where are they based?
• How long is input stored, and how is it deleted after the contract ends?

The answers belong in writing in your contract records. When something happens – for example an enquiry from the Austrian Data Protection Authority (dsb.gv.at) – what counts is what you can prove.

What if company data has already gone into an AI tool?

The realistic finding in many companies: it was tried out long ago – with real quotes, real customer emails, real cost calculations. A sober four-point plan helps here:

1. Take stock: which tools were used, with which accounts and roughly with which data? Ask without assigning blame – otherwise you will learn nothing.
2. Correct the settings immediately: deactivate use for training wherever there is a switch; delete stored histories as far as possible.
3. Move to an orderly setup: introduce a business version with a data processing agreement (DPA) and training exclusion, stop using private accounts for company purposes.
4. Request deletion: apply to the provider for deletion of stored input – the GDPR gives data subjects and controllers tools for this. To be honest: what has actually gone into a model can hardly be removed afterwards. That makes the rule for the future all the more important.

What does this mean in practice for your company?

The checkpoints result in a simple company rule: company data only goes into tools for which there is a data processing agreement (DPA) and a contractual training exclusion – and only via the official company accounts. This rule should be put in writing and communicated to staff, otherwise shadow IT fills the gap.

For an AI knowledge base that holds your company’s collected knowledge, the standard is particularly strict: instructions, service records and experience are the company’s capital. A reputable provider excludes use for training by contract, keeps the data on EU servers and backs every AI answer with the source document – so every statement remains verifiable, because even a well-connected AI can be wrong.

The training exclusion is one building block among several – alongside server location, data processing agreement (DPA) and access control. The overall picture, including a provider checklist, is given in the article AI and the GDPR in Austria.

As a next step: take the two or three AI tools already in use in your company today and go through the five checkpoints – this rarely takes longer than an hour per provider. How the introduction of a knowledge base with a contractual training exclusion works is shown on the page on the process of the introduction.