Metilda Clinical Insights
AI Receptionist UK GDPR Compliance: What the Data Use and Access Act Actually Requires
Your AI receptionist's vendor page probably says it's GDPR compliant. Here's what that badge doesn't tell you: the law it describes changed on 5 February 2026, and most Cliniko-connected vendors still displaying it haven't updated the sentence since.
Take Aeva AI, an established Cliniko-connected AI receptionist for allied health clinics. Its own Cliniko marketplace listing states that its compliance "also meets the New Zealand Privacy Act and the UK & EU GDPR." That claim is live on its listing page today, and it was almost certainly accurate when someone wrote it. It is no longer a complete description of UK law. On 5 February 2026, the UK's Data (Use and Access) Act 2025 (the DUAA) repealed Article 22 of the UK GDPR, the provision that governs automated decision-making, which is what an AI receptionist does every time it books, reschedules, or triages a call with no human reading the transcript first, and replaced it with four new articles: 22A, 22B, 22C and 22D. A generic "meets the UK & EU GDPR" sentence says nothing about whether a vendor has adjusted to that. The pattern isn't unique to Aeva: Solium AI's listing claims it is "Compliant with the HIPAA, GDPR, and the Australia Privacy Act," and Evie Assist AI's listing claims compliance with "HIPAA, GDPR, POPIA, Australia and New Zealand privacy acts." Neither mentions the DUAA or Articles 22A-22D. None of this means these vendors are doing anything unlawful. It means the badge predates the law it cites, and a clinic can't tell from the badge alone whether anyone has gone back and checked.
That gap matters more for an allied health clinic than for most AI-receptionist buyers, because the calls themselves routinely involve special-category data (a disclosed condition, an injury, a mental health concern, a reason for cancelling), and the new UK regime treats that category differently from everything else. This post covers what actually changed, what Articles 22A-22D require of a calling clinic specifically, and a worksheet for classifying your own clinic's call types rather than taking a listing page's word for it.
The UK GDPR Badge Most AI Receptionist Vendors Still Use
Before 5 February 2026, Article 22 of the UK GDPR ran on a default of prohibition. A decision made solely by automated means, with a legal or similarly significant effect on a person, wasn't allowed unless it was necessary for a contract, authorised by law, or the person had given explicit consent. That's the regime most "GDPR compliant" marketing copy in this category was written against.
The Data (Use and Access) Act 2025 received royal assent on 19 June 2025. Section 80 and Schedule 6 replaced Article 22 with Articles 22A to 22D, and commencement regulations brought that change into force on 5 February 2026, confirmed in the UK's own legislation.gov.uk text of Schedule 6, which rewrites the UK GDPR's cross-references to point at the new articles, and in Bratby Law's commencement-date breakdown, which lists "Articles 22A to 22D UK GDPR (recast automated decision-making)" against 5 February 2026 under commencement regulation SI 2026/82. The default flipped: under the new Article 22C, a significant automated decision is now permitted provided the controller applies safeguards, rather than prohibited unless an exception applies.
Five dates, one open question: the operative legal change already happened on 5 February 2026, but the regulator's own explanation of how to apply it is still a draft seven months later.
The ICO opened a public consultation on its draft automated decision-making guidance on 31 March 2026, alongside a separate report on recruitment-sector ADM use, and the consultation closed on 29 May 2026, both dates confirmed on the ICO's own consultation page. As of a UK data-protection commentary last updated 16 October 2026, the final guidance still hadn't been published; the piece describes the ICO's position as "draft guidance" with the final version expected "before the code," referring to a separate statutory AI/ADM code of practice the ICO is developing in parallel. The law changed on a fixed date seven months ago; the regulator's own worked explanation of how to apply it is still unsettled. That's an awkward position for a clinic wanting a straight answer, and part of why a vendor's compliance badge isn't a substitute for working through the test yourself.
Automated Decision-Making Under the New Articles 22A-22D
The four new articles don't carry equal weight for a clinic evaluating a calling AI. Article 22A supplies the definitions that make the other three operable. Article 22D fills in supplementary detail, including rule-making power for further regulations. The two that actually determine what a clinic has to do sit in the middle:
- Article 22C sets the new general default. A significant automated decision is permitted, provided the controller gives the person meaningful information about it, lets them make representations, provides a real opportunity for human intervention, and lets them contest the outcome. Wiggin's summary of the provisions frames this as the core shift: old default prohibition, new default permission subject to safeguards.
- Article 22B is the carve-out keeping the old, stricter rule alive for one category: decisions relying on special-category data, which includes health information. Brodies' analysis of the ICO consultation summarises its effect as keeping the prohibition on significant automated decisions built on special-category data, lifted only with explicit consent, or where processing is necessary for a contract or required by law and also meets a substantial-public-interest condition under Schedule 1 of the Data Protection Act 2018.
For an AI receptionist taking allied-health calls, Article 22C mostly governs ordinary bookings. Article 22B is the one that matters the moment a caller says something about their health: the one provision in this reform that didn't get easier.
The Three-Question Test for Automated Decision-Making AI Receptionist Calls
Strip out the legal language, and the ICO's draft guidance, as summarised by the firms tracking it (the final text isn't out yet), reduces to a sequence of three questions, asked in order, for any given call outcome:
The test is sequential, not a checklist you can score all at once: a "yes" at the human-involvement question ends the analysis before the other two questions are even relevant.
The ICO's draft guidance (still a draft) defines a decision as "solely automated" when there's no meaningful human involvement in it; a person skimming a transcript after the fact, or rubber-stamping whatever the AI already did, doesn't count. "Significant effect" covers anything that meaningfully affects someone's access to a service, their finances, or their legal position, so denying or confirming a medical appointment plausibly qualifies, given it affects access to care. The third question is genuinely new for this sector: did the decision rely on data the caller disclosed about their health, mental health, or similar.
Where UK Health Data Rules Diverge From Australia's APP 1.8
If you've already worked through Metilda's earlier post on Australia's APP 1.8 disclosure obligations for AI receptionists, it's worth being precise about why Article 22B isn't just the UK's version of the same thing. APP 1.8 is a disclosure duty: it requires a practice to be transparent, generally in its privacy policy, about how it uses automated decision-making and handles cross-border data. It doesn't gate the processing on a specific additional consent before it happens; it's a "tell people what you're doing" obligation, enforced through transparency.
Article 22B isn't satisfied by disclosure alone. Where a significant automated decision is built on special-category data, the processing itself is prohibited unless the controller has the person's explicit consent for that specific processing, or can point to a substantial-public-interest condition under Schedule 1 of the Data Protection Act 2018. A privacy policy that accurately describes what the AI does is necessary either way, but in the UK it isn't sufficient on its own for the special-category branch of this test, which is stricter, and more UK-specific, than the comparable Australian duty.
Same fact pattern, two different legal mechanisms: the UK path requires an affirmative basis before the processing happens; the Australian path requires honest disclosure about processing that's already happening.
None of the three listings checked for this post (Aeva AI's, Solium AI's, Evie Assist AI's) describe collecting explicit, decision-specific consent before acting on a disclosed health condition, or relying on a specific Schedule 1 public-interest ground. That doesn't mean none of them do it; a marketing listing isn't where most vendors would document consent-capture mechanics. It does mean a UK clinic can't verify the Article 22B question from the listing page alone, which is exactly the gap a generic "meets the UK GDPR" badge leaves open.
Classify Your Clinic's Call Types: A GDPR Compliance Worksheet
Run each of your clinic's actual call outcomes through the same three questions, in order. Here's how six common allied-health call types land, reasoned through explicitly so you can substitute your own clinic's specifics where they differ from the assumptions stated.
Autonomous booking (the AI books a new appointment against the live Cliniko calendar, no human review). Human involvement: no. Significant effect: plausibly yes, since it determines whether the caller gets a slot. Special-category data: no, assuming the caller disclosed nothing beyond the appointment type itself. Result: Article 22C; the four safeguards apply, but it isn't blocked outright.
Retry/rebooking (the first attempt fails because a slot disappeared, and the AI retries automatically). Same reasoning as autonomous booking: no human involvement, a plausible significant effect on access to care, no special-category data in the clean case. Result: Article 22C.
FAQ answering (hours, fees, parking, what to bring to a first appointment). Human involvement: no. Significant effect: no, since nothing about the caller's access, finances, or legal position is being decided; the AI is relaying static information. Result: outside Article 22's scope entirely, because no decision about the individual is being made.
Human-assisted handoff (the AI gathers information; a staff member reviews it and makes the actual booking or triage call). Human involvement: yes, provided the review is genuine rather than a rubber stamp. Result: not "solely automated," so the Articles 22A-22D analysis doesn't apply here at all. Ordinary UK GDPR processing rules still apply, but the ADM-specific test doesn't.
Triage/escalation flagging (the caller mentions a symptom, say chest pain or a mental health crisis, and the AI's own assessment decides whether the call escalates immediately, with no human checking that flag). Human involvement: no. Significant effect: yes, since it affects how quickly the caller reaches care. Special-category data: yes. Result: Article 22B applies; the clinic needs the caller's explicit consent to that specific processing, or a genuine Schedule 1 public-interest ground, not just a general privacy-policy disclosure.
Auto-applied fee or cancellation decision (the AI hears the caller's stated reason for cancelling and decides whether a late-cancellation fee applies, no human review). Human involvement: no. Significant effect: yes, it's financial. Special-category data: depends what the caller said. "Something came up" keeps this in Article 22C territory; "I was admitted to hospital" pulls it into Article 22B, because the fee decision now rests on disclosed health information. This is the call type most likely to drift between regimes within the same call flow, which is exactly why a blanket compliance badge can't answer the question for a whole product category at once.
The pattern across all six: the first two questions are usually answerable from the call type alone, but the third, special-category data, depends on what a specific caller actually says, not on the product's design in the abstract. That's harder for a vendor to certify once and be done with, and a reasonable part of why none of the listings reviewed here go further than a generic GDPR mention.
Frequently Asked Questions
Does my AI receptionist count as "automated decision-making" under UK law?
Only for call outcomes where there's no meaningful human review and the outcome has a legal or similarly significant effect, not for every call it handles. An AI answering a question about opening hours isn't deciding anything about the caller, so that sits outside Article 22's scope, while one that autonomously books or declines an appointment is making exactly the kind of decision Articles 22A-22D were written to cover.
My vendor's page says it's GDPR compliant. Isn't that enough?
A generic GDPR statement tells you the vendor engages with UK and EU data protection law in some form, but not whether they've reviewed their product against the automated-decision-making provisions that changed on 5 February 2026. None of the three Cliniko-connected listings checked for this piece mention the Data (Use and Access) Act or Articles 22A-22D by name, so there's no way to confirm from the listing alone whether that review has happened.
What actually changed on 5 February 2026?
Article 22 of the UK GDPR, which previously prohibited significant automated decisions by default, was replaced by Articles 22A-22D. The new default under Article 22C permits significant automated decisions as long as the controller applies four safeguards. Article 22B kept the old, stricter prohibition in place specifically for decisions built on special-category data, including health information.
Does a caller mentioning a health condition during a booking call trigger something special?
It can. If that disclosure drives a significant, solely-automated decision (a triage priority, an escalation, a fee waiver), Article 22B requires explicit consent for that processing, or a substantial-public-interest ground under Schedule 1 of the Data Protection Act 2018. If a human reviews the decision before it's finalised, the Article 22B question doesn't arise, because the decision is no longer "solely automated."
Is this the same obligation as the Australian rule I've already looked at?
No. Metilda's earlier post on Australia's APP 1.8 disclosure duty (linked in the previous section) covers a transparency requirement: telling people, generally in a privacy policy, how automated decision-making and cross-border data handling work. Article 22B is different, since it can prohibit the processing itself unless a specific legal basis (explicit consent or a public-interest ground) is in place, not just disclosed.
When will the ICO's final guidance be published, and should I wait for it before doing anything?
As of a UK data-protection commentary last updated 16 October 2026, the final guidance still hadn't been published; the consultation closed 29 May 2026, and the ICO has said final guidance is expected before a related statutory AI/ADM code of practice, with no committed date. Waiting for the final text before working out which of your own call types fall under Article 22B isn't a safe default. The underlying law has been in force since 5 February 2026 regardless of when the explanatory guidance catches up.
What UK Clinics Should Actually Do Before the ICO's Final Guidance Lands
The final guidance isn't out yet, and there's no committed date for it. That's a reason to work through the parts of this that don't depend on the guidance text, not a reason to wait.
- Run the three-question test against your own call types now, using the worksheet above as a starting template rather than the six labels verbatim, since your clinic's actual call flows may split differently, especially around triage and fee decisions.
- Ask your vendor a specific question, not a general one. Instead of "are you GDPR compliant," ask whether they've reviewed their product against Articles 22A-22D, and whether any call type can rely on disclosed health information to reach a significant decision without human review. A vendor's answer (or non-answer) is more informative than any badge on a listing page.
- For any call type landing in Article 22B territory, check what legal basis actually applies: explicit consent captured for that processing, or a genuine Schedule 1 public-interest ground, rather than assuming a general privacy-policy mention covers it.
- Update your own patient-facing privacy notice to reference the current law. A policy still citing "Article 22" without the 22A-22D framing is a smaller version of the same gap this post opened with.
- Set a reminder to re-check this once the ICO's final guidance is published, since the final text could adjust how "meaningful human involvement" or "significant effect" gets interpreted in practice, even though the underlying law isn't expected to change again.
One disclosure, in the interest of the standard this post is holding vendors to: Metilda hasn't published an independent audit against Articles 22A-22D either. The design intent behind its human-assisted handoff and escalation flows is to keep triage decisions that touch disclosed health information on the human-involvement side of this test rather than the solely-automated side. That's a design goal being built toward, not a certification already held.
Sources
- Aeva AI — Cliniko Connected Apps listing
- Solium AI — Cliniko Connected Apps listing
- Evie Assist AI — Cliniko Connected Apps listing
- Data (Use and Access) Act 2025, Schedule 6 — legislation.gov.uk
- Bratby Law — Data (Use and Access) Act 2025: changes now in force
- Wiggin — Key provisions of the Data (Use and Access) Act 2025 come into force
- ICO — Consultation on the draft guidance about automated decision-making, including profiling
- Brodies — Automated decision-making: ICO consults on updated guidance
- NeuralTrust — GDPR and AI: a UK CISO's guide (updated 16 October 2026)