What the DPC says about
AI, and what to do about it
The Data Protection Commission’s 2024 guidance on AI and large language models asks deployers to understand where personal data goes, set retention, assess risk and support rights requests. Its September 2026 AI Insights Report adds lawful basis and transparency findings. This page turns both into a checklist for an AI agent.
What has the DPC published on AI?
The Data Protection Commission (DPC) is Ireland’s GDPR regulator. It has also been a market surveillance authority under the EU AI Act since 2025, and in July 2026 its AI Act role was extended to several more areas, including some biometric uses and most prohibited practices. Three publications matter most for a business that deploys an AI agent:
| Publication | Date | Why it matters to a deployer |
|---|---|---|
| DPC blog: AI, Large Language Models and Data Protection | 18 July 2024 | Written for organisations that build AI and for those that use third-party AI tools. The basis of the checklist below. |
| EDPB Opinion 28/2024 on AI models | December 2024 | Requested by the DPC. Sets EU criteria for legitimate interests, anonymity of models and the effect of unlawful training. |
| DPC AI Insights Report: Responsible AI Innovation (2021 to 2025) | 25 September 2026 | Findings from DPC work on about 180 AI products and services. The DPC says it also serves as guidance for controllers. |
The DPC has not published guidance specific to AI agents that we could find. The 2024 blog addresses organisations that use AI products supplied by a third party, which is the position of most businesses with an agent.
The DPC’s main points for organisations that use AI
- Know where the data goes. Before you adopt an AI product, understand what personal data it uses, where the data goes when a third party is involved, and whether the provider keeps or reuses it. The provider’s documentation should tell you this clearly.
- Staff inputs create risk. Personal data that staff type into an AI tool can lead to processing you did not plan.
- You may be a controller. If so, consider a formal risk assessment. A DPIA may be needed, in particular for new technology.
- Set retention. Without a retention schedule and processes to support it, you risk breaching storage limitation.
- Support rights requests. Put processes in place so people can exercise their rights, including access and erasure, for data held in the AI system.
- Training creates further processing. Personal data used to train or fine-tune a model can lead to unwanted processing, and models can reproduce passages of their training data.
- Outputs can be wrong. Inaccurate or biased outputs used without critical human review can create risks, including from automated decision-making.
- Filters can fail. Filters meant to block personal data can sometimes be bypassed. Ask how the provider deals with this.
The 2026 AI Insights Report focuses on two issues: the use of legitimate interests as a legal basis for AI training, and the need for strong transparency for a new technology. It also gives particular attention to children. It says that early engagement with the regulator is the most effective path for a new product.
A DPC-based checklist for an AI agent
Use this before launch. Each line maps to a point in the DPC guidance or the GDPR article behind it.
| Area | What to do | What to keep as evidence |
|---|---|---|
| Lawful basis | Name a basis for each purpose. Contract for handling a customer’s own booking or order; legitimate interests, with a written balancing test, for triage or summaries. | Purpose and basis table; balancing test |
| Transparency | Update the privacy notice to explain the agent, the processors and any transfers. Say at the start of each conversation that the person is dealing with an AI. | Notice version and date; screenshot of the first message |
| DPIA | Screen every agent. Do a full DPIA for health data, children, monitoring or large-scale processing. Do it before launch. | Screening note; DPIA |
| Processors | Sign a DPA with the model provider, messaging platform, hosting provider and the developer who runs the agent. | Processor register with DPA references |
| Transfers | Record where prompts, outputs, logs and backups are processed and stored. Use EU processing where it is available and practical. | Transfer map; mechanism for each transfer |
| Retention | Set periods for transcripts, tool logs, summaries and search indexes. Match the model provider’s retention to your notice. | Retention schedule; deletion job evidence |
| Rights requests | Make sure you can find, export and delete one person’s data across the agent, the CRM and the provider. | Tested access and erasure procedure |
| Training on customer data | Turn off any provider setting that uses your data to train its models. Do not fine-tune on customer conversations without a basis and a DPIA. | Screenshot of the setting; written confirmation |
| Human review | Keep a person in any decision with legal or similar effect on someone. Check samples of agent replies each week at first. | Escalation rules; review log |
Our DPIA template for AI agents follows this structure, and our guide to GDPR and AI agents explains each area in more depth.
Training on customer data and the EDPB opinion
Most business agents do not train a model. They send each conversation to a provider’s model and get a reply. The training question then is about the provider: does it use your inputs to improve its models? Most business and API terms say no by default, but the setting is per account. Check it for the exact account the agent uses. Our page on EU data residency for AI models lists the defaults for the main providers.
If you want to fine-tune a model on your own customer conversations, the EDPB opinion and the DPC report become directly relevant. You need a lawful basis for that new purpose. If you rely on legitimate interests, people must be able to object. Tell them before the training happens, and remove or pseudonymise identifiers where you can. In many cases, retrieval instead of fine-tuning gives the same result with less personal data in the model.
Where agent projects usually fall short
- Logs nobody planned for. The agent framework keeps full transcripts in a debugging tool with no retention limit.
- Free-text special category data. Patients and clients mention health, finances or family matters without being asked. Plan for it, as our clinics page explains.
- No route for objections. Under legitimate interests, a person can object. The agent needs a way to stop processing and pass the person to staff.
- Notice updated after launch. The privacy notice should change before the first real conversation, not after a complaint.
If your DPIA finds a high risk that you cannot reduce, Article 36 of GDPR requires you to consult the DPC before you start the processing. This is rare for a well-scoped business agent. It is more likely for large-scale health data or for profiling that affects people’s access to services. The DPC’s 2026 report encourages early engagement for new AI products, so if you are building something novel, contact the DPC early rather than after launch.
We build these controls into every AI agent we deliver: field-level data access, set retention, deletion by phone number or email, and a first-message disclosure.
Common questions
Has the DPC published guidance on AI agents?
Not specifically on agents, as far as we could find on 19 November 2026. Its July 2024 blog on AI, large language models and data protection covers organisations that use third-party AI products, which includes most agent deployments. Its September 2026 AI Insights Report adds findings on lawful basis, transparency and children.
Do I need a DPIA for a customer service agent?
Not always. A DPIA is mandatory where processing is likely to result in high risk, for example large-scale special category data or systematic monitoring. A simple booking agent may not meet that test, but the DPC blog says a risk assessment should be considered when you act as controller. Record your screening decision either way.
Can we use customer conversations to improve our agent?
You can review conversations to fix prompts and rules, if your privacy notice covers it and you keep to your retention periods. Using them to train or fine-tune a model is a new purpose. It needs its own lawful basis, clear notice before training and, under legitimate interests, a way for people to object.
Is the DPC also an AI Act regulator?
Yes. Irish designation regulations name the DPC as a market surveillance authority for several areas, including some biometric uses, law enforcement, migration and justice, and most prohibited practices. GDPR complaints about AI systems also go to the DPC, as before.
Sources
- DPC: AI, Large Language Models and Data Protection (18 July 2024)
- DPC: Data Protection Commission publishes AI Insights Report (25 September 2026)
- DPC: AI Insights Report, Responsible AI Innovation (PDF)
- DPC: Data Protection Impact Assessments
- EDPB: Annual Report 2024 (Opinion 28/2024 on AI models)
- Irish Statute Book: S.I. No. 405/2026 (DPC designations under the AI Act)
We last checked these sources on 19 November 2026. Rules, fees and prices change. Check the source before you act.
Bring us the task.
Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.
Start a project.

Book a call or send us a message. We read everything.
tash@neu.ie
Working with teams in Ireland and worldwide.
Free 30-minute AI review.