An AI agent can process customer data, but the answer is not automatically yes. It depends on the purpose, types of data, legal basis, security and the role of each supplier. The GDPR applies whenever personal data is processed. Specific transparency duties under the EU AI Act also apply from 2 August 2026. The practical route is to map exactly which data the agent touches, why that is necessary, where it is processed and which human control fits the risk.
This is manageable when arranged up front. Research from the Dutch Chamber of Commerce in 2025 found that three in ten business owners use ChatGPT or Copilot, but only 7 percent say they are well informed about the AI Act. That gap between use and knowledge is where risks can arise.
What does the GDPR say about an agent?
There is no separate exemption, and that is the point. The Dutch Data Protection Authority explicitly identifies AI agents as applications that can process personal data. The existing GDPR questions therefore apply here too:
- Legal basis: why are you allowed to process this data, and does that basis also cover processing by an agent?
- Purpose limitation: customer data you collected to send invoices can鈥檛 quietly become training fuel or analysis material for something entirely different.
- Roles and processors: if an external platform processes the data on your behalf, define the roles correctly and put a data processing agreement in place where required.
- Transfers: does the data leave the European Economic Area, or can subprocessors outside it access the data? If so, check the transfer mechanism and supplementary safeguards.
Those last two points make the architecture choice at least as important as the model choice. Local or European processing can limit international transfers, but does not make an application GDPR-compliant by itself. Legal basis, purpose limitation, security, retention, data-subject rights and supplier arrangements still apply. These are also the points Renforza wants to map before setting up agents, executed together with our technical sister company aitonomi.
What changes on August 2, 2026?
From that date, the transparency duties in Article 50 of the AI Act apply. The European Commission summarises the core requirement as informing people when they interact with an AI system, unless that is obvious from the circumstances. Certain AI-generated or manipulated content is also subject to specific marking or disclosure duties. This does not mean every internal AI draft needs the same visible label; the duty depends on the role, context and type of content.
For a customer-facing chatbot, a clear and timely notice is generally the logical implementation. For generated content, first determine which specific duty applies. The difficulty lies not only in the interface, but also in correctly scoping the application. Anyone putting an agent live this summer should complete that analysis before launch.
Apart from the AI Act, there is another reason to work transparently and with controls: liability. In Moffatt v. Air Canada, the airline was held responsible for incorrect information from its chatbot. The German Higher Regional Court of Hamm attributed misleading chatbot statements to the operator, as summarised by the Library of Congress. The legal basis differs by country and case, but a business cannot simply hide behind the software.
The check in five questions
Want to know whether your planned agent passes inspection? Run through these five questions:
- Which personal data does the agent touch? Not roughly, but precisely: which fields, from which systems.
- Where does that data get processed? EU, locally, or via a US service? Keep asking until you have a concrete answer, including about subprocessors.
- Is there a data processing agreement with every party in the chain?
- Which AI Act transparency duty applies here? For direct chatbot interaction, informing the user is required in principle from 2 August 2026 unless the AI nature is obvious. Different, more specific rules apply to content.
- Which risk assessment and human control are needed? Money, contracts, profiling, special-category data or decisions with major effects may call for additional review, potentially including a DPIA. Why human control matters in practice is covered in our piece on what an AI agent can鈥檛 do.
If you can answer all five concretely, you have a useful first inventory. That is not legal approval, but it is the basis for a focused privacy and AI Act assessment.
Is this a reason to wait?
Not necessarily. The rules do point in the same direction as sensible design: scoped, with known data sources, appropriate security and human control at critical moments. That does not make compliance free, but it prevents privacy and transparency from being added only after the build.
My take: treat the AI Act the way you wish you鈥檇 treated the GDPR in 2017. Not as a reason to panic, but as a reason to do it right from the start. In an agents intake we run these five questions as standard, alongside the question of whether the process itself is agent-suitable. That way you know before the build begins whether your customer data can take part, and under what conditions. For the broader employer view, see our page for employers.
Sources and legal scope
- European Commission: AI Act and application timeline
- European Commission: transparency duties from 2 August 2026
- Dutch Data Protection Authority: GDPR conditions for generative AI
- Dutch Chamber of Commerce: business awareness and preparation for the AI Act
This article is a practical first inventory, not legal advice. Specialist review is sensible for special-category data, profiling, automated decision-making or major effects on individuals.
