Back to blog

AI support and GDPR: data processing agreements and what to check

Published on October 21, 2026

Two Humane figurines sign a contract at a bright table, a small padlock beside it.

A support email is rarely just a question. It contains a name, an email address, often a delivery address, an order number, sometimes a phone number, and occasionally things the customer didn't need to write at all: “I'm in hospital right now, so I can't accept the parcel.” If you hand that email to an AI service so it can reply, you're processing personal data through a third party. The GDPR has clear rules for that, and they apply to a shop with two people just as they do to a corporation.

According to the Bitkom study 2025, 53 percent of companies name legal uncertainty and 48 percent strict data protection requirements as obstacles to using AI. That's understandable, but most of it can be settled with a handful of questions you can put to any vendor before signing a contract.

This article explains which roles you and the vendor take on, what a data processing agreement is and what it must contain, which points to check beyond it, and what the EU AI Act additionally requires from 2026. It stays general. This is not legal advice.

Who is who: controller and processor

The GDPR has two roles that matter to you. You, the shop, are the controller: you decide that customer data is processed and for what purpose. The AI vendor that reads your emails and writes replies is usually the processor: it processes the data on your behalf and according to your instructions.

That has two consequences. First: the responsibility stays with you. If the vendor loses data or uses it for other purposes, you're the one the supervisory authority and the customer turn to. Second: you may only give data to a processor if a contract exists that governs the processing. That's the data processing agreement (in German: Auftragsverarbeitungsvertrag, AVV), regulated in Art. 28 GDPR.

No data processing agreement, no data transfer. It's the simplest rule in this article and the one most often overlooked, usually because a tool is “just being tried out” and is already reading real customer emails while that happens.

What the data processing agreement must contain

Art. 28(3) GDPR lists what a processing agreement must cover at minimum. Translated into questions you can put to the contract:

  1. Does the vendor process only on your instructions? Concretely: it uses your customer emails to answer your requests, and for nothing else.
  2. Are the vendor's staff bound to confidentiality?
  3. Which technical and organisational measures are in place? Encryption, access control, logging. This is the reference to Art. 32 GDPR.
  4. Which sub-processors are involved? An AI vendor almost always uses further services, for hosting or the language model, for instance. They must be named, and you must be informed about changes.
  5. Does the vendor support you with data subject requests? If a customer requests access or deletion, you must be able to cover the data held by the vendor as well.
  6. What happens to the data when the contract ends? Deletion or return, with a deadline.
  7. Can you verify that the vendor complies? Right to information, evidence, in case of doubt an audit.

A reputable vendor has the data processing agreement as a standard document you can read and sign without involving a lawyer. At Humane you'll find it under Data processing agreement. If a vendor can't produce one on request, or says “that's not necessary with us”, the conversation is over.

The questions beyond the agreement

The data processing agreement is the obligation. The following points decide whether you sign it with a clear conscience.

Where is the data stored?

Under Art. 44 GDPR, a transfer to a third country may only take place under certain conditions. For you the practical question is: on which servers do the emails sit, and where does the language model run? The two can be in different places. Ask about both. At Humane, data is stored on German servers and encrypted with AES-256.

Are your customer conversations used for training?

This is the question many forget. Some vendors use their customers' conversations to improve their models, for all customers. Then your customer's email about her hospital stay ends up in a training dataset. In most cases that's processing your customer doesn't expect and you have no basis for. Ask explicitly: is my data used to train models for other customers? At Humane the answer is no. The agent learns from your corrections, but only for your shop.

Who on the vendor's team sees your emails?

The vendor's support staff, developers during debugging, people with admin access. A vendor should be able to explain who has access and when. And on your side: who on your team has access to the dashboard? If you bring in temporary staff or freelancers, they should only see what they need. How that works with permissions per task is in Helpers instead of full-time.

What happens in a data breach?

If something happens at the vendor, under Art. 33 GDPR you may have to notify the supervisory authority within 72 hours. That's only possible if the vendor informs you without delay. It should be in the agreement; ask anyway what the process looks like in practice.

Your own obligations

The data processing agreement governs the relationship with the vendor. Three things remain with you regardless.

  • Duty to inform. Under Art. 13 GDPR you must inform customers how their data is processed. In practice: the AI vendor belongs in your privacy policy, as a recipient of the data, with the purpose “handling support requests”.
  • Record of processing activities. If you keep one, AI-assisted support handling is added as an entry.
  • Automated decisions. Art. 22 GDPR restricts decisions made solely by automated means that have legal effect. An automatic reply to “Where is my parcel?” isn't such a decision. An automatic rejection of a refund could go in that direction. That's another reason why refunds and rejections should go through approval by a human.

How high the fines can theoretically be is in Art. 83 GDPR: up to 20 million euros or 4 percent of worldwide annual turnover, whichever is higher. For a small shop that isn't the realistic order of magnitude. Realistic is a customer's complaint to the supervisory authority, followed by an enquiry you have to answer. For that you want the data processing agreement, the privacy policy and an answer to “where is the data” ready to hand.

What the EU AI Act requires from 2026

Alongside the GDPR, Article 50 of the EU AI Act applies from 2 August 2026. The core: AI systems that interact directly with people must be designed so that the people concerned know they're interacting with an AI, unless that's obvious from the circumstances.

For support this means: if an AI replies to customers, that should be recognisable. Exactly how the disclosure must look in email traffic hasn't been settled in every detail yet. In practice, a short, honest note, in the signature for instance, and a clearly recognisable path to a human work well. Neither harms the customer relationship. Customers react far worse to hidden automation than to open automation.

The checklist for vetting a vendor

Before you hand real customer emails to an AI service:

  • Data processing agreement under Art. 28 GDPR in place, read, signed.
  • Server location for data and model known.
  • Commitment that your conversations aren't used to train models for other customers.
  • List of sub-processors received.
  • Encryption and access control described.
  • Process for a data breach clarified.
  • Deletion after the end of the contract settled.
  • Vendor listed in your privacy policy.
  • Dashboard permissions for your team set so everyone sees only what they need.
  • Note on AI use towards customers prepared.

What the agent needs to know about your shop in terms of content, so it answers correctly, is a different question, and only marginally about data protection: What an AI agent needs to know about your shop. And how to stop it claiming things that aren't true is in Hallucinations in customer service.

In short

AI in support isn't a special case for data protection; it's ordinary commissioned processing. You're the controller, the vendor is the processor, and between you there must be a data processing agreement. Beyond that, three questions count: where the data sits, whether it's used to train for others, and who can see it. List the vendor in your privacy policy, keep refunds and rejections behind a human approval, and make the use of AI recognisable to customers. The bigger picture is in AI in customer service: what actually works in 2026. This is not legal advice.

If you want to check how we've set this up: try 100 customer requests for free, no credit card, with the data processing agreement and privacy policy available before the first ticket.

Sources