# AI agents and GDPR: where does the email data live?

Putting an AI agent on an inbox creates a new flow of personal data. Here's what that means under GDPR — and the questions your DPO will ask first.

When your team adds an AI agent to an inbox, you have — in GDPR terms — created a new processing activity. Personal data that was already sensitive (names, contact details, the contents of correspondence, sometimes special-category data) now moves somewhere it didn't before: to an LLM, usually through a SaaS platform, often across a border. GDPR does not treat that as "just an AI feature." It is a data flow, and as the controller you are accountable for it.

The first question a competent DPO will ask is the one in the title. Here is how to answer it — and the related questions that follow. They are vendor-neutral; they hold whoever builds the agent. (Disclosure: we build Mailbuttons. We've kept this piece honest, so it's useful regardless of what you choose.)

# 1. You've created a new processing activity — write it down

An agent reading and acting on email is processing personal data. Your accountability obligations apply to it like any other system.

Ask:

• Has this processing been added to your Article 30 record of processing activities — data categories, purpose, legal basis, recipients, retention?
• Is the purpose compatible with why the email was originally collected? Triaging support mail usually is; a novel use may need its own basis.
• Does it warrant a DPIA? Systematic, large-scale processing of correspondence by an automated system often does.

# 2. Follow the email — every hop is a location

"Where does the data live" is really several questions, one per hop. The message lands in a mailbox, an agent platform reads it, and its contents are sent to a model. Each hop is a place and a processor.

Ask:

• Where are the mailboxes hosted?
• Where does the agent platform itself run?
• Which model provider receives the message content, and in which jurisdiction is it processed?

# 3. International transfers — the model is usually the catch

Mailbox hosting can be EU/UK; the model is where it often slips. Most frontier LLM APIs process in the US. Sending EU or UK personal data there is a restricted transfer — since Schrems II it needs a lawful basis: an adequacy decision, Standard Contractual Clauses, or the UK IDTA, with supplementary measures where required.

Ask:

• Does any hop leave the EU/UK? If so, which transfer mechanism covers it, and is it documented?
• Can you constrain the processing path to your required region if your risk appetite demands it?

# 4. The processor chain, your DPA, and model training

The agent platform is your processor under Article 28 — that requires a DPA. It, in turn, relies on sub-processors: the model provider, hosting, storage. You are entitled to know who they are.

Ask:

• Is there a signed DPA with full Article 28 terms?
• Is there a current sub-processor list, and will you be notified before it changes?
• Does the contract prohibit the model provider from training on your data?

# 5. Minimise what the model sees

Data minimisation is not optional. Not every email needs an LLM: a vendor invoice or a SaaS notification can be delivered and retained without a model ever reading it. A platform that lets you scope which mail reaches the agent is giving you a minimisation control, not just a convenience.

Ask:

• Can you configure which senders or message types the agent processes, versus those it merely delivers?
• Is there a way to keep mail on file without sending it to the model at all?

# 6. Records, retention, and data-subject requests

Accountability means being able to show what was processed. Retention limits and erasure apply to the email and to whatever the agent records. And a subject access request will now reach the agent's processing too.

Ask:

• Is there a log of what the agent processed and did — one you could rely on as a record?
• Is retention configurable, for both mail and logs, to match your policy?
• Could you locate, export, or erase one data subject's data across the mailbox and the logs?

# The short answer

"Where does the email data live" should have an answer you can give in a sentence: the country the mailbox sits in, the country the model runs in, and who the processors are. If a platform cannot tell you that plainly, that is your finding.

We built Mailbuttons EU/UK-hosted, with a DPA, and a policy layer that lets you decide which mail ever reaches the model — so the question has a short, documented answer rather than a shrug. mailbuttons.com