Voice AI Privacy Notices: The Article 13 Transparency Guide

On a website, the privacy notice is a solved problem. You put a link in the footer, a banner at the top, and a layered notice behind a "learn more" click. On a phone call handled by a voice agent, none of that exists. There is no footer, no banner, no screen to point at. Yet the legal duty is identical: when a UK controller collects personal data directly from a person, it must tell that person who is processing their data, why, on what legal basis, for how long, and what rights they have. That duty is UK GDPR Article 13, and it applies to the first sentence of an automated call just as much as to a web form.

This is the transparency problem almost no voice AI buyer has actually solved. Most enterprises deploy an agent that says "this call may be recorded" and assume the compliance box is ticked. It is not. Recording disclosure is one small fragment of a much longer list, and in 2026 that list changed: the Data (Use and Access) Act 2025 amended Article 13 itself, adding a new required element and rewiring two others. Adoption is racing ahead of this. McKinsey's State of AI, published in November 2025, found that 88% of organisations now use AI in at least one function while only 6% capture material earnings impact, and the transparency layer is exactly the kind of unglamorous control that gets skipped in the rush to ship.

This guide is shipped by the team behind Dilr Voice, enterprise voice AI built for regulated deployments. Or see DATS, our five-stage AI consulting system.

What is a voice AI privacy notice under Article 13?

A voice AI privacy notice is the fair processing information a controller must give a caller when a voice agent collects their personal data. Under UK GDPR Article 13, Dilr Voice and every other platform must convey the controller's identity, the purposes and legal basis, the recipients, the retention period, and the caller's rights. On an audio channel this cannot be one spoken paragraph, so it becomes a layered notice.

The right to be informed sits underneath the whole of data protection. It is the mechanism that makes every other right usable: a person cannot exercise access, objection or erasure over processing they were never told about. That is why the ICO's right to be informed guidance treats the "privacy notice" as shorthand for a family of methods rather than a single document. Enterprises running voice programmes should read it alongside our voice AI data retention guide, because the retention period is one of the elements the notice has to state.

The reason this matters commercially is that transparency failures are cheap to prevent and expensive to remediate. A missing element is designed into every call your agent makes, so a single drafting gap becomes thousands of identical infringements before anyone notices. Getting the notice right once, at design time, is the whole game, which is why it belongs in an AI operating model rather than a post-launch scramble.

What information must a voice AI privacy notice include?

Article 13 sets out a fixed list of elements, and in 2026 it grew. A compliant voice AI privacy notice must cover who the controller is, the purposes and legal basis, the recipients, any international transfers, the retention period, the caller's rights, and any automated decision-making. The Data (Use and Access) Act 2025 added one more: the right to complain to the controller.

The single hardest constraint is not the content, it is the timing. The statute is explicit about when the information has to land.

"Where personal data relating to a data subject are collected from the data subject, the controller shall, at the time when personal data are obtained, provide the data subject with all of the following information."

That is UK GDPR Article 13(1), and the phrase "at the time when personal data are obtained" is what makes a phone call difficult. You cannot save the notice for the end of the call, because by then the data is already collected. The table below maps the current required elements and how each one behaves on a voice channel.

Two of these are 2026 changes that most published guidance has not caught up with. The Data (Use and Access) Act 2025 amended Article 13 to insert a new element 13(2)(ca), the right to complain to the controller under section 164A of the Data Protection Act 2018, which came into force on 19 June 2026. Separately, from 5 February 2026, the automated decision-making element at 13(2)(f) now cross-refers to the safeguards required by Article 22C rather than the repealed Article 22, so any voice agent that makes a solely automated decision needs to reflect that, and our voice AI automated decision-making guide covers the underlying duty. The transfer element at 13(1)(f) also changed on that date, which our international data transfer guide tracks against the new UK rules.

When must you give a voice AI privacy notice on a call?

You must give it at the point of collection, not afterwards. Article 13 requires the controller to provide the information at the time when personal data are obtained, so on an inbound call the opening layer has to play before the agent captures details, and on an outbound call it comes immediately after identity is established. Dilr Voice treats the opening disclosure as a gated step: the agent cannot capture data until the first layer has played.

Timing has two further wrinkles the statute now spells out. If you later want to use the data for a genuinely new purpose, Article 13(3) requires you to tell the person about that other purpose before the further processing starts, which is why model training on recordings is a transparency question and not just a security one. Our purpose limitation guide for call recordings works through that scenario in detail. The one relief valve is Article 13(4): you do not have to repeat information the person demonstrably already has, though on a first-time inbound call you can rarely lean on that.

How do you deliver an Article 13 notice when there is no screen?

You deliver it in layers, because you cannot say twelve elements aloud without the caller hanging up. The ICO endorses this: it advises against restricting privacy information to a single notice, and suggests using a device's audio to provide key information "through device speakers, complemented by more detailed information available in a written notice." For Dilr Voice that means a short spoken layer, a just-in-time line, and a clear signpost to the full notice.

The layered pattern is the practical answer to the audio constraint. The spoken layer carries the elements a reasonable person needs to hear in the moment: who is calling, why, and that the call is handled by an AI agent and may be recorded. The full notice, which the caller can read at leisure, carries the exhaustive list. A just-in-time notice sits between them, triggered by context: before the agent asks for a card number or a health detail, it says one plain sentence about what that specific data is used for. This is the surface where a governed platform earns its keep, and it is a core part of how we think about an AI operating model for voice.

The fourth node is the one enterprises forget. Because the notice is spoken, there is no static page to point an auditor at. You have to be able to prove which version of the notice played on which call, which means logging a notice version identifier against each call record. That evidencing layer is where a build-it-yourself stack and a governed platform diverge most sharply, and it is worth pressure-testing on your voice AI agents before you scale traffic.

How is Article 13 different from the EU AI Act Article 50 duty?

They are two separate duties on the same call. Article 13 of UK GDPR is a data protection duty: the controller must explain who is processing the caller's personal data and why. Article 50 of the EU AI Act is a transparency duty: a person must be told they are interacting with an AI system unless that is obvious. A voice agent can trigger both, so buyers who conflate them satisfy one and miss the other.

The practical consequence is that "we tell people it is a bot" does not discharge Article 13, and "we have a privacy notice" does not discharge Article 50. Dilr Voice treats them as distinct fields in the disclosure step. Our dedicated EU AI Act Article 50 disclosure guide covers the AI-disclosure duty in full, including where the "obvious" exemption applies. Enterprises selling into both the UK and the EU should map both duties against every call type rather than assuming one notice covers everything, which is how we approach dual-regime deployments at Dilr.ai.

Who owns the privacy notice, the enterprise or the voice AI vendor?

The controller owns it, and the controller is usually the enterprise, not the vendor. Article 13 places the duty on the controller, which decides why and how personal data is processed. A voice AI platform such as Dilr Voice, or a build-your-own stack on Vapi, is usually a processor: it makes delivery and logging possible but does not decide what the notice says. Getting this allocation wrong is a live enforcement risk the vendor cannot cure.

This is where the privacy notice connects to the wider contractual picture. If the roles are misallocated in your data processing agreement, the notice duty lands in the wrong place, and our Article 28 controller-versus-processor guide works through how to fix that. The vendor's real job is enablement: exposing the notice text as a configurable, versioned asset, playing the right layer at the right moment, and writing the notice version into the call log so the controller can evidence compliance in an ICO audit. That is the standard we hold ourselves to across the DATS methodology, and it is a fair question to put to any vendor before signing.

What is the best way to handle voice AI privacy notices in 2026?

The best approach depends on how much notice logic you want to own. If you run a single call type with strong in-house data protection engineering, a build-your-own stack on Vapi, Retell AI or Bland AI can work, at the cost of maintaining it as Article 13 changes. If you run many call types across regulated lines, a governed platform such as Dilr Voice or PolyAI is the stronger fit, because notice layers and evidencing arrive as managed features.

There is no single winner, only a fit to your risk profile and your appetite for maintenance. The concession is real: a small team automating one low-risk booking line will often move faster on a raw builder platform than on a governed one, and that is a legitimate choice. What is not legitimate is treating the notice as an afterthought on any platform. The 2026 changes, the new complaint-to-controller element and the Article 22C cross-reference, mean a notice drafted in 2025 is already out of date, so whichever platform you choose, the notice needs an owner and a review cadence. If you want a second opinion on your specific call estate, our execution office runs exactly this kind of review.

Want to see this in production? Try Dilr Voice live, book an AI placement diagnostic, see our DATS methodology, or read about our approach to placing AI inside regulated call flows.

Frequently asked questions

Do I need consent to record the call, or just a privacy notice?

These are different obligations. A privacy notice under Article 13 tells the caller what you do with their data; consent is one possible legal basis for the processing, and often not the basis you rely on. Many voice AI calls run on legitimate interests or contract rather than consent, so you still owe the full notice but do not need a yes. Our consent capture guide for voice calls shows which basis fits which call.

Does the research exemption let me skip the notice for model training?

Almost never. The Data (Use and Access) Act 2025 inserted Article 13(5) to (7), which can switch off the further-processing notice, but only where the new purpose is genuine scientific or historical research, archiving in the public interest or statistical purposes, done in line with Article 84B, and only where giving the notice is impossible or a disproportionate effort. Training a commercial voice model on recordings rarely meets that bar, so the transparency duty usually still bites.

What if the notice is buried in a long spoken script?

Burying it fails the transparency principle. Article 5(1)(a) of UK GDPR requires processing to be transparent, and the ICO expects privacy information to be concise and intelligible, not smuggled into a monologue nobody follows. On a voice channel the spoken layer must be short and clear, with the detail moved to the signposted full notice. A notice the caller cannot absorb is treated as no notice at all, so brevity here is a compliance feature, not a nicety.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.