What Regulators Are Actually Saying About AI in Customer Service

When a support agent pastes a customer's details into a chatbot to draft a reply, it feels like using a smarter spell-checker. Regulators see something else: a controller sending personal data to a third party. Over 2024 and 2025, three of Europe's most influential authorities made their expectations explicit. None of them are waiting for the EU AI Act to start enforcing.

CNIL: no DPIA, no deployment

France's CNIL has been among the most active regulators on AI. Its published AI guidance and “how-to” sheets are consistent on one point: where an AI system processes personal data in a way that is likely to be high-risk, a Data Protection Impact Assessment is the starting point, not an afterthought. Running customer records through a general-purpose model at scale is exactly the kind of processing that triggers it.

CNIL has also been explicit in its messaging to organisations: using a tool like ChatGPT for customer service without a Data Processing Agreement and a documented risk assessment is a compliance failure. The DPA establishes the Article 28 processor relationship; the DPIA shows you assessed the risk before exposing people's data. Skip either and you have processing the regulator considers unlawful by default.

ICO: AI does not dilute accountability

The UK's Information Commissioner's Office takes a complementary line. Its guidance on AI and data protection makes clear that AI-assisted customer service must satisfy the same data-protection standards as any human-handled processing. There is no lighter regime because a model is in the loop.

The accountability principle is the crux: it is not enough to be compliant, you must be able to demonstrate it. Introducing AI adds to what you must evidence (the lawful basis, the safeguards, the human oversight). It does not subtract from it.

In practice this reverses a comfortable assumption. Teams often treat AI as a productivity add-on that sits outside the compliance perimeter. The ICO's position pulls it firmly inside: the moment customer data reaches the tool, every existing obligation applies, and the burden of proof is on you.

EDPB: customer-facing AI is on the radar

At EU level, the European Data Protection Board coordinates how national authorities approach emerging risks. It convened a dedicated ChatGPT taskforce and runs a coordinated enforcement framework that sets shared priorities across member states. The direction of travel is unambiguous: generative AI in customer-facing rolesis precisely the category supervisory authorities are organising around for 2026.

That matters because coordinated priorities turn into coordinated action. When the EDPB signals a focus area, multiple national regulators investigate the same theme in parallel, and from 2 August 2026 they can examine the same incident through the EU AI Act as well as the GDPR.

What the three positions add up to

Read together, the guidance converges on four practical expectations:

• Assess before you deploy. A DPIA for AI processing of personal data is the baseline, per CNIL.
• Paper the processor relationship. No DPA means no lawful route for customer data into the tool.
• Hold the same bar. AI-assisted handling meets the same standard as human handling, per the ICO, and you must be able to show it.
• Expect scrutiny. Customer-facing generative AI is a stated enforcement focus, per the EDPB's direction for 2026.

The awkward gap in most organisations is the last mile: even with a policy, a DPIA, and an approved enterprise tool, employees still paste customer data into their personal ChatGPT tab. Policy describes the intent; it does not produce the evidence a regulator asks for.

From policy to provable control

Each regulator is, in effect, asking the same question: can you show what happens when customer data meets an AI tool? Answering it requires a technical control at the point of exposure (the browser), plus a record you can hand to your DPO.

1. Prevent the unlawful path. Block pastes of customer personal data into unsanctioned AI tools on-device, before the data leaves the browser. The disclosure that never happens needs no breach analysis.
2. Evidence the safeguard. Maintain an exportable, timestamped log of detections (classification, user, tool, severity) as the accountability evidence the ICO and CNIL expect. Record classifications, never the content itself.
3. Feed the DPIA. Real usage data turns a DPIA from a paper exercise into a living document: which tools, which data, how often, trending which way.

The question to put to your team

“If a regulator asked today how we stop customer data reaching consumer AI tools, and to show the evidence, what would we hand them?”

If the answer is a policy document and a hopeful shrug, you are exposed on exactly the points CNIL, the ICO, and the EDPB have already spelled out. The fix is not a committee. It is a control at the browser and a log for your DPO.

This article summarises regulators' published positions for general information and is not legal advice. See our compliance overview for the framework.