What processing actually takes place?
Start here, because without this answer you cannot fill in the rest.
As soon as an AI model processes personal data, this constitutes processing within the meaning of the GDPR. This applies to a customer name in a CRM record just as much as to an email containing a complaint.
Map out: which categories of personal data pass through, for what purpose, and on what legal basis. For most SME applications, this is legitimate interest or the performance of a contract, but you must be able to substantiate this.
What you have after that is the basis for your processing register. You must maintain that in any case.
What belongs in the data processing agreement?
Your AI supplier is a processor. This means a data processing agreement, which must include at a minimum: what is processed and for what purpose, how long data is retained, which sub-processors are engaged, where the processing takes place, which security measures apply, and what happens upon termination.
Specifically check two points that differ with AI vendors compared to standard software: whether your data is used to train models, and how long prompts and responses are retained for abuse detection, for example.
| Ask | Short answer | Floor |
|---|---|---|
| What processing actually takes place? | Start here, because without this answer you cannot fill in the rest. | Read more |
| What belongs in the data processing agreement? | Your AI supplier is a processor. | Read more |
| How do you handle retention periods? | An AI connection creates copies that you do not always see. | Read more |
| What does data classification mean in practice? | Not all data needs to be able to pass through an AI connection. | Read more |
| Who is responsible? | You, as the data controller. | Read more |
How do you handle retention periods?
An AI integration creates copies that you don't always see. A summary of a customer file that ends up as a note in your CRM is a new process with its own retention period.
Agree on where output lands and how long it remains. In practice, the simplest rule is: AI output follows the retention period of the record to which it belongs. If the customer file is cleaned up, the summary is included.
Also pay attention to logging. Logs that record who requested what contain personal data themselves and therefore also require a time limit.
What does data classification mean in practice?
Not all data needs to be able to pass through an AI connection. Data classification sounds daunting, but it comes down to three categories.
What can always be included: standard customer and process data that your employees already have access to.
What is only allowed under certain conditions: more sensitive material such as contracts, financial data, and personnel information. Here, you restrict access to specific roles.
What may never be included: special personal data such as health data, and anything for which you have no legal basis.
You translate that classification into scopes in your integration. That is the moment when policy becomes technology.
Who is responsible?
You, as the data controller. That does not change because you engage a supplier.
In practical terms, this means you must be able to demonstrate which choices you have made and why. A short document stating which data is accessible, to whom, on what legal basis, and for what retention period is usually sufficient for this. That is also what you produce when a customer or regulator asks for it.
Where do you start?
First create an overview of which data can pass through the connection before configuring the connection. It works the other way around as well, but then you do the work twice.
Sources and product information
- Zoho Security FAQ, official product information
- Anthropic Privacy Center, official product information