What GDPR actually requires from any company deploying AI agents

Most teams shipping an AI agent in Europe treat GDPR as a checkbox at the end of the project: add a cookie banner, write a privacy policy, done. That's not what the regulation asks for, and it's not what gets a company in trouble when something goes wrong. GDPR is a design constraint, and it applies in full to agents, not just to forms and databases.

An agent is a "processing activity" the moment it touches personal data

If an AI agent reads a customer's name, reads a patient's chart, or even infers something about a person from behavioral data, GDPR's obligations apply to that agent the same way they'd apply to any other system: lawful basis for processing, purpose limitation, data minimization, and a documented record of what's happening and why.

The common mistake is assuming an LLM-based agent is somehow exempt because "it's just reading text, not storing it in a database." Processing includes reading, not just storage. If the agent's underlying model call sends personal data to a third-party API, that's a data transfer, and it needs the same lawful basis and safeguards as any other transfer.

The requirements that actually bite

• Lawful basis. You need a specific legal ground — consent, contract necessity, legitimate interest — for the agent's processing, and it has to match what the agent is actually doing, not a generic basis picked at project kickoff.
• Data minimization. An agent with a knowledge base built from "everything we have" instead of "everything this task needs" is already out of compliance, independent of how secure the storage is.
• Right to explanation and human review. Article 22 gives individuals the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects, without meaningful human involvement. An agent that auto-approves or auto-denies something consequential — a loan, a claim, an eligibility decision — needs a human review path, not just good intentions.
• 72-hour breach notification. If an agent's access gets compromised and personal data is exposed, the clock on notifying the relevant supervisory authority starts at discovery, not at confirmation. Logging what an agent touched is what makes that 72-hour window survivable instead of a guessing exercise.
• International transfers. If the model provider processes data outside the EU, you need a valid transfer mechanism (Standard Contractual Clauses, an adequacy decision, or equivalent) — this is exactly why running an agent against a private, in-region deployment rather than an arbitrary third-party API matters in practice, not just in theory.

What this means for how you actually build

None of this requires avoiding AI agents. It requires building them the way any other system touching personal data should be built: scoped access, logged actions, minimized data, and a way to explain a decision when someone asks. The teams that get burned aren't the ones using AI — they're the ones who treated "it's an AI feature" as a reason the usual rules didn't apply.