How does GDPR apply to an AI agent in an Irish business?
An AI agent that reads messages, customer records or documents is processing personal data, so GDPR applies in full. Your business is usually the controller. You need a lawful basis, contracts with every processor, a plan for data leaving the EU, sensible retention and clear notice that people are dealing with an AI.
Where does GDPR touch an AI agent?
An AI agent rarely works with anonymous data. It reads a customer’s WhatsApp message, looks up their booking, drafts a reply and writes a note to your CRM. Every step is processing personal data. A language model does not change the rules; it changes where the data goes.
In most setups your business is the controller: you decide why the data is processed and what the agent does with it. The company that hosts the model, the messaging platform, your hosting provider and whoever builds and runs the agent are usually processors acting on your instructions. In Ireland, the regulator is the Data Protection Commission (DPC), which publishes guidance for organisations on its website.
Which lawful basis, and how little data does the agent need?
Every purpose the agent serves needs a lawful basis under Article 6 of GDPR. For most business agents the realistic options are contract, legitimate interests and consent:
- Contract: answering a customer about their own booking, order or account, because you need to process their data to deliver the service they asked for.
- Legitimate interests: for example, triaging inbound enquiries or summarising emails for staff. This needs a written balancing test showing your interest is not overridden by the person’s rights.
- Consent: for optional uses, such as marketing follow-ups. Consent must be freely given and as easy to withdraw as to give.
Health information, and other special category data, needs an additional condition under Article 9. A clinic using an agent for bookings should assume patients will mention symptoms in free text. See our notes for clinics.
Data minimisation is where agent design matters most. Give the agent the fields it needs for the task, not the whole customer record. If it confirms appointments, it needs a name, a time and a service, not payment history or clinical notes. Avoid loading entire mailboxes or shared drives into a knowledge base “in case”. Our scoping guide builds this into the first step.
Who are your processors, and what do the contracts say?
Article 28 requires a written contract, usually called a data processing agreement (DPA), with each processor. It must cover what they process, on whose instructions, security measures, sub-processors, help with data subject requests, breach notification and what happens to the data at the end. Most large providers publish a standard DPA; check it covers the account the agent uses.
| Party | What to check |
|---|---|
| AI model provider | Whether your inputs are used to train models, how long prompts and outputs are retained, and which business or API terms apply. See [OpenAI](/integrations/openai/) and [Anthropic](/integrations/anthropic/). |
| Messaging platform | For WhatsApp, Meta processes message content and metadata under its business terms. Read them alongside your privacy notice. |
| Hosting and database | The region where data is stored, backups, and who at the provider can access it. |
| Your developer | If they run the agent for you, they are a processor too and need a DPA, access controls and an offboarding plan. |
| Connected systems | Your CRM, calendar or accounting tool already has terms with you, but the agent may now move data between them. |
Keep a simple register: each processor, what data they receive, the DPA reference and the region.
A GDPR checklist for your AI agent.
Before an agent handles personal data
A working checklist, not legal advice. Your DPO or solicitor should review the result.
0 of 10 done
What happens when data leaves the EU?
Many model providers are US companies. A transfer of personal data outside the European Economic Area needs a legal mechanism under Chapter V of GDPR. The common ones are:
- An adequacy decision. The EU-US Data Privacy Framework covers transfers to US companies that have certified under it. Check the provider’s certification yourself rather than relying on a sales page, and note that the framework has faced legal challenges, so keep an eye on its status.
- Standard contractual clauses, usually built into the provider’s DPA, with a transfer risk assessment on your side.
- EU data residency, where a provider offers EU-hosted processing for your account or plan. This reduces the transfer question but does not remove the need for a DPA.
Ask each provider in writing where prompts, outputs, logs and backups are stored and processed. Logs are the part people forget.
When do you need a DPIA, and how long should you keep conversations?
A data protection impact assessment is mandatory where processing is likely to result in a high risk to people’s rights and freedoms. The DPC’s guidance gives examples such as large-scale profiling, large-scale processing of special category data and systematic monitoring, and it publishes a list of processing types that need one. Processing that meets two or more of the European high-risk criteria will generally need one. An agent that screens job applicants or handles patient messages at volume is a likely candidate.
Do the DPIA before the agent goes live and keep it under review. It should describe the processing and its purpose, assess necessity and proportionality, set out the risks to people, and record the measures you take. If a high risk remains after those measures, GDPR requires you to consult the DPC before going ahead.
On retention, decide periods for each kind of data the agent creates: conversation transcripts, tool call logs, drafts, summaries and any embeddings in a search index. Logs are needed for an audit trail, but keeping every transcript forever is hard to justify. Delete or anonymise on a schedule, and confirm the model provider’s own retention matches what your privacy notice says.
Where an agent would make decisions with legal or similarly significant effects on people without a person involved, Article 22 adds further rules. Keeping a person in those decisions is the safer design.
Do you have to tell people they are talking to an AI?
Under GDPR, your privacy notice must explain what you do with personal data, including the use of an AI system, the processors involved and any transfers. Update it before launch. In the conversation itself, say at the start that the person is using an automated assistant and tell them how to reach a person.
The EU AI Act adds its own transparency duties in Article 50. In summary, AI systems that interact directly with people must be designed so those people know they are dealing with an AI, unless that is obvious from the context. There are separate duties for marking AI-generated content and for disclosing deepfakes and certain AI-generated text published on matters of public interest. These duties apply from 2 August 2026, and parts of Article 50 have since been amended through the EU’s “digital omnibus” changes, including timing for the content-marking duty. Check the current text, or ask your advisor, for how it applies to your setup.
An agent that opens with a clear disclosure and offers a route to a person is in a good position under both. Neu builds that disclosure into every customer-facing agent, including WhatsApp agents.
A short checklist before the agent goes live
- Write down each purpose the agent serves and its lawful basis.
- List the data fields it can read and write, and remove anything it does not need.
- Confirm a DPA with every processor and record where each stores data.
- Decide whether a DPIA is needed, and keep the reasoning either way.
- Set retention periods for transcripts, logs and indexes, and automate deletion.
- Update your privacy notice and add an AI disclosure at the start of conversations.
- Check how you will answer access and erasure requests that include agent data.
- Review access and logging with our AI agent security guide.
Common questions
Will the AI provider train its models on our customers’ data?
It depends on the provider, the product and your account settings. Business and API offerings from major providers generally state that inputs are not used for training by default, while some consumer products differ. Check the current terms for the exact account the agent uses, confirm the setting in writing, and record it in your processor register.
Do we need a data protection officer to use an AI agent?
Not automatically. GDPR requires a DPO for public bodies and for organisations whose core activities involve large-scale monitoring or large-scale special category data. Many small businesses do not meet that test. Even so, someone should own data protection for the project, and outside advice is worth getting if the agent handles sensitive data.
How do we handle an access or erasure request that covers agent data?
Treat agent transcripts, logs and CRM notes like any other personal data. Your system should let you find a person’s records by phone number or email, export them and delete them, including copies held by the hosting and model providers where they retain data. Design this before launch, because retrofitting search across logs is harder.
Is an AI agent riskier under GDPR than ordinary software?
The principles are the same, but agents tend to receive free text, which brings in data you did not ask for, and they often send it to more third parties. That widens the exposure. Narrow scopes, field-level access, short retention and clear disclosure bring the risk back to the level of a normal business system.
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.