Skip to main content
Data & security

Is Your SaaS Tool Sending Your Company's Data Out to Train AI Models? How to Find Out

Many SaaS tools have had an AI feature added without customers being asked whether their data may be used to train the models. Here is how to find the specific answer for your own tools.

Author: Agent360 editorial team
Published: 23 July 2026
Last updated: 7 November 2026
Reading time: 13 min read

How do you find out whether your SaaS sends data to AI training?

The answer lies in three specific documents: the data processing agreement (DPA), the vendor's list of sub-processors, and special AI terms. Look for words such as "product improvement", "model training" or "aggregated data", and for whether there is an opt-out.

You can't see it in the user interface, and you can rarely work it out by asking a salesperson. The answer lies in three specific documents: the data processing agreement (DPA), the vendor's list of sub-processors, and the special AI terms that many vendors have added after launching an AI feature. Look for wording about "product improvement", "model training" or "aggregated data", and for whether there is an opt-out. If all three are unclear or absent, that is a red flag, and you should ask the vendor directly, in writing, before you carry on using the feature on customer data.

Why the question is coming up now

A great many of the SaaS tools Danish companies already use (CRM, support chat, meeting platforms, marketing automation) have had an AI feature added in recent times. Sometimes as a visible new button, other times quietly in the background as part of an update. The problem is that the terms you once clicked "yes" to don't necessarily cover the new use of your data.

That isn't malicious on the vendors' part. It is a consequence of product development moving faster than contract revision. But the consequence for you as a customer is the same: you don't know for certain whether messages from your customers, notes from your salespeople or minutes from your meetings are being used to train a model that other customers (or the public) can potentially benefit from later.

It isn't a question you can answer by "having a feeling about it". It requires you to go in and read three places in your vendor relationship.

What AI training on your data actually means

When a vendor writes that data may be used "to improve our services" or "for model training", it typically means that the inputs and outputs from your use of the AI feature (messages, documents, transcriptions) become part of the dataset on which the vendor's, or an underlying AI provider's, model is trained or fine-tuned.

That isn't necessarily the same as your specific customer data being "leaked" to other customers. But it does mean the data leaves the context in which you originally provided it, and is used for a purpose you may never have explicitly accepted. And if your customers have shared personal data with you (which they almost always have, in a CRM or a support queue), the decision is no longer yours alone to make. It is a decision that affects your customers' data.

This is where it becomes relevant to know the difference between a data processor relationship and one in which the vendor genuinely acts as a data controller for its own purposes (such as product improvement). GDPR requires the roles to be clearly defined, and any use of personal data for new purposes to have a legal basis. If you are in doubt about your own role as data controller, the Danish Data Protection Agency's guidance is a good place to start; see the list of sources at the bottom.

The three places you need to look

1. The data processing agreement (DPA)

The data processing agreement is the legally binding document that must (among other things) describe the purposes for which the vendor may process your data, and whether there are restrictions on further use. A well-written DPA will explicitly rule out the use of customer data for model training, or require separate consent if it is to happen.

The problem is that many DPAs were signed before the vendor had an AI feature. So even though the DPA formally still applies, it may say nothing about AI training, because it wasn't written for that scenario. That doesn't mean you are covered: it means the question is unresolved, and that you should ask the vendor for an updated annex or a written confirmation.

2. The sub-processor list

Few SaaS vendors build their own AI model from scratch. Most pass data on to an underlying model provider, which could be a large language model vendor or a cloud-based AI service. This relationship must appear on the sub-processor list, which under GDPR the vendor is obliged to make available and to update when new sub-processors are added.

Check specifically: Is there an AI model provider on the list that wasn't there a year ago? Were you told about the addition (as a starting point you should have the opportunity to object to new sub-processors)? And does it state where the data is processed geographically: within the EU/EEA, or outside it, where a special transfer basis is required?

3. AI-specific terms or a "trust center"

Several vendors now have a separate AI terms document or a "trust" page where they describe how the AI feature handles data in practice. This is typically where you find the most up-to-date (and most honest) description of whether data is used for training, whether there is an opt-out, and whether the opt-out applies retroactively or only going forward.

Find this document and read it literally. Vague wording such as "we may anonymise and use aggregated insights" is not the same as a guarantee that your raw data isn't included in training.

What to check, where to find it, and what counts as a red flag

What to check Where in the agreement Red flag
Purpose of data processing Data processing agreement (DPA), main text The DPA says nothing about the AI feature; signed before it was launched
Use for model training AI terms / trust page Wording about "product improvement" without clarifying whether it includes model training
Underlying AI model provider Sub-processor list A new provider added without you being informed
Option to opt out Account settings / AI terms No visible opt-out, or an opt-out only for new data, not data already submitted
Geographic location of data Sub-processor list / DPA annex Transfer outside the EU/EEA with no transfer basis stated
Separation from other customers' data AI terms / security documentation No description of whether training takes place across customers or in isolation

Typical vendor answers, and what they really mean

When you ask the questions above, you rarely get a clean yes or no back. You get wording that sounds reassuring but needs translating. Here are some of the most common, and what you should listen for when you read them.

"We only use anonymised or aggregated data" sounds safe, but is only a guarantee if the vendor can also explain how the anonymisation actually happens, and at what point in the process it takes place. If the anonymisation happens after the raw data has already been sent to an external model provider, it hasn't prevented your raw data from leaving your control; it has only changed what is left behind afterwards.

"Data is only used to improve your experience of the product" is wording that can cover anything from simple bug fixing in the code to your messages becoming part of the training data for a model that other customers also benefit from. Ask for clarification: does "improve the experience" specifically mean model training, or does it mean something narrower, such as error logging and performance measurement?

"We follow industry standards for data security" says something about the level of security (encryption, access control and the like) but nothing about what the data is used for. It's an answer to a different question from the one you asked, and shouldn't be confused with an answer about the purpose of the data processing.

"It's described in our terms" is often true, but the terms may be fifteen pages long and written so that it takes a lawyer to decode them. Instead, ask the vendor to point to the specific section that answers your question: if they can't do that quickly, it is in itself a sign that the question hasn't been asked very often before.

Who in the organisation should own the question

It is tempting to let the question drift between IT, procurement and the department that actually uses the tool, without anyone ending up owning it. In practice it's a good idea to place responsibility in one place: typically with whoever already has responsibility for data processing agreements and GDPR compliance in the company, whether that is a dedicated compliance function, part of IT operations, or an external adviser that smaller companies draw on.

This person or function should also be the one to approve new AI features before they are rolled out to the whole organisation, not just the employee who first spots the button in the user interface and thinks it looks useful. A simple internal rule can be: no new AI feature in an existing SaaS tool is activated for customer data until the person responsible has seen and approved the AI terms in writing.

Why your own instance changes the calculation

Part of the uncertainty described above arises because most SaaS tools are built as one shared platform, where your data lives side by side with other customers' data in the same infrastructure, and where it is the vendor's decision, not yours, whether and how AI features process that data.

We are currently building Agent360 OS, in which each customer can have its own instance and its own database; it hasn't been released yet. That doesn't change the fact that you still need to read your agreements carefully and understand what happens to data: no vendor, whatever the architecture, relieves you of that responsibility. The idea is that the question "will our data be mixed with other customers' data, or used to train a model that other customers benefit from" gets a structurally different starting point when data by its very nature isn't shared across customers in the same instance.

How to ask the vendor specifically

Don't just call or write "do you use our data for AI?". Instead, put these questions in writing, so that you have the answer on paper:

  1. Is data we enter in [the specific AI feature] used to train or fine-tune your own or a third party's AI models?
  2. If yes: is there an opt-out, and does it also apply to data we have already entered?
  3. Which underlying AI model provider processes our data, and does it appear on your sub-processor list?
  4. Is our data processed within the EU/EEA? If not, which transfer basis is used?
  5. Is there a difference in how you treat data on our paid plan versus a free/trial plan?

The fifth question isn't random. Several vendors have historically applied more lenient terms to free users than to paying customers, and if you have created test accounts yourselves or use free tiers for parts of the organisation, it's worth checking whether the terms are the same.

What the AI should NOT do

An AI agent or an internal tool shouldn't itself decide whether your data may be used for external model training, and it shouldn't interpret the law for you. That is a decision that requires a person who has read the specific agreement and who, in cases of doubt, brings in a lawyer or the Danish Data Protection Agency's guidance. An AI agent can help gather the documents, flag unclear wording and keep track of which vendors are missing answers, but it doesn't make the final assessment.

How to keep track going forward

It isn't a one-off check. Vendors update terms and AI features on an ongoing basis, and a new feature can come with new data processing terms without you being actively asked to accept them again. So set up a fixed routine:

  • Review your critical SaaS vendors' AI terms and sub-processor lists at regular intervals, e.g. once a quarter.
  • Appoint one person internally (typically part of the IT or compliance remit) who receives and assesses notices of new sub-processors.
  • Before you activate a new AI feature in an existing tool, check the AI terms first; don't just activate it because the feature appears in the user interface.
  • Document your assessment in writing, so that you can show it if a customer or the Danish Data Protection Agency asks.

Frequently asked questions

Q: Is it illegal for a vendor to use our data for AI training?

Not automatically. It depends on whether there is a valid basis for processing, whether it is described correctly in the agreements, and whether the purpose is consistent with what you originally told your customers. It is a concrete assessment that should be made together with a lawyer or by following the Danish Data Protection Agency's guidance.

Q: Can we simply say no to our data being used for AI training?

It depends on the vendor's terms. Some offer an explicit opt-out in the account settings, others require you to contact support, and some give no option at all for certain plan types. It is one of the things you need to clarify directly with the vendor.

Q: Does the same apply to free versions and paid subscriptions?

Not necessarily. Several vendors have historically had different data processing terms for free/trial users versus paying customers. Check the terms for the specific plan you use.

Q: What if the vendor doesn't answer our questions clearly?

An unclear or evasive answer is itself important information. Document that you have asked, and consider whether it is a risk you can accept, or whether it should feed into a decision to switch vendor or solution for the workflow concerned.

Q: Is your own instance the same as a guarantee that data will never be used for AI training?

No. Your own instance and your own database change the structural risk of data being mixed with other customers' data or ending up in a shared training pool, but it is not a legal guarantee, and it doesn't relieve you of reading your agreements. This is not legal advice.

Q: Where do we find the Danish Data Protection Agency's official guidance on data processors and AI?

The Danish Data Protection Agency's full rules and guidance are available on its website; see the link in the list of sources below. If you are in doubt about a specific situation, it is best to seek legal advice or contact the Danish Data Protection Agency directly.

This article is not legal advice. Contact a lawyer or the Danish Data Protection Agency for an assessment of your specific situation.

Sources:

#Data and security #GDPR #AI governance #SaaS #data processing agreement

Unsure about the rules for your own set-up?

We look at what you use today and what it requires: documentation, consent and what the EU AI Act means for you. 20 minutes, no commitment.