How We Use AI

Our AI providers

Our main AI provider is Anthropic PBC (United States), the company behind the Claude models, under its Data Processing Addendum with the UK International Data Transfer Agreement. What Anthropic's published terms say:

• Anthropic's Commercial Terms of Service say that Anthropic may not train models on customer content from its services — which covers what we send it through its API.
• Anthropic's published retention policy for commercial customers says it deletes inputs and outputs within 30 days, unless they are flagged under its usage policy (then kept for up to two years) or the law requires otherwise.

We also use two further AI providers, both named in our Privacy Policy:

• Google (Gemini) — the default model for AI chat, which receives the parts of your record you select with every message. Gemini also reads DUTCH hormone reports, recognises a document's type and laboratory, and reads a sample of Randox reports so that we can compare its quality with Claude's (that answer is not used). Google's Gemini API terms say that, for customers in the UK, Google does not use prompts or responses to improve its products, and keeps logs for a limited period to detect misuse.
• Deepgram — speech-to-text for optional voice entry. The short audio clip you record is sent to Deepgram, and the text comes back to the field you would otherwise type into; we do not keep the audio. Our requests ask Deepgram not to keep the audio; under its terms, audio from such requests is kept only as long as needed to transcribe it.

A third provider, Perplexity AI, Inc. (United States), can be reached only through a clinician compare option described in the table below, which appears only if a Perplexity key is configured on our server.

Your record itself stays on our own server in the EEA ( Germany). For each AI call, the data sent to the provider is the data listed for that feature in the table below.

Where a laboratory printed its own comment, flag or reference range beside a blood result on a report you uploaded, those words are stored on your record verbatim, and the AI chat sees them as part of your record whenever your question is about that marker. When you ask about a marker, its comment, printed range and flag are passed to the model quoted and attributed to the named lab and report date. In the general list of your results the comment itself is not passed at all — the list says only that a comment exists and points you to that marker — and a printed flag travels with its marker name and test date rather than the lab's name. No AI feature is ever asked to summarise, shorten or reword a lab comment, and nothing on your record — no reminder, task or diagnosis — is ever created from one.

Which models we use

Which model handles which job is set in our configuration; this list shows the setting on the date at the top of this page. The specific model used for each AI response is recorded in the internal audit log. You can request a copy of your audit log entries at any time.

Every AI feature in Zaynix Health

This list is updated with each release (last checked against the code on 6 October 2026). If a feature is not on it, it does not use AI; if you find one that does, tell us at privacy@zaynix.health and we will correct the list. Rows that start “Clinician” are tools used by clinicians at a clinic you are connected to.

What we never do

• We never train AI models on your data ourselves. Whether a provider may use what we send is set by that provider's terms: Anthropic's commercial terms do not allow it to train on our content; Google's Gemini API terms say it does not use prompts or responses from UK customers to improve its products; our voice-entry requests opt out of Deepgram's model-improvement programme, so under its terms that audio is kept only as long as needed to transcribe it (see “Our AI providers” above).
• We never use AI to make decisions about you that have legal or similarly significant effects (UK GDPR Article 22). What AI produces is shown to a person — you, or a clinician you are connected to — who decides what, if anything, to do with it. Values AI reads from your documents, and entries it drafts from what you type or say, are proposals: they reach your record only when you confirm them (the Patient Confirmation Layer).
• We never use AI to describe or interpret your record without your consent. Summaries and descriptions, the import-review questions, the allergy check's AI step, explaining a medicine when you tap “What is this for?” and every clinician AI tool run only if AI processing is switched on in Settings → Privacy, and AI chat asks for its own, separate consent instead (it does not read the AI processing setting). You can switch AI processing off at any time. Some features use AI to do a job you ask for — reading a document you upload, drafting an entry from what you type or say, transcribing voice entry and recognising a medicine's name — and these run when you use them.
• We never hide that AI is in use. Every AI feature is listed on this page with the data it sees, and a link to this page sits at the foot of every page of the web app for patients. AI summaries of blood markers carry a small purple “AI” badge that opens this page; other AI features label their output in words — for example “AI-generated” or “AI Analysis” — and AI chat tells you that it is an AI. A relationship the allergy check's AI step found says so (“An AI lookup links …”), and a medicine explanation is labelled “AI-generated”. Not every AI output is labelled yet.
• We never sell your data to anyone, AI provider or otherwise.

What we always do

• Record every AI call in an internal audit log, with the model, the feature, token counts and a fingerprint of what was sent. You have a right to a copy of your audit log.
• Mark every AI-generated record in your GDPR data export with an ai_generated flag so you can tell which parts of your data came from AI.
• Keep your record usable without AI interpretation: switch AI processing off and your record, charts and timeline keep working, and you can type entries in by hand instead of uploading documents.
• Keep a person in charge: AI output is shown to you, or to a clinician you are connected to, to read and judge. It is not a substitute for clinical judgement.

Risks and limitations you should know about

Large language models like Claude are powerful but imperfect. Zaynix Health manages this by:

• Writing nothing that AI reads or drafts to your record until you confirm it (the Patient Confirmation Layer).
• Cross-validating extracted drug names against NHSBSA dm+d (licensed medicines) and clinical concepts against SNOMED CT UK, via the NHS Terminology Server. Lab-marker codes are being rebuilt and are shown only where the code has been verified.
• Rate-limiting AI calls per user per month so an errant prompt injection cannot drain your quota or ours.
• Running a dedicated ai_audit_log that records every call for post-hoc review.

Your controls

• Grant or revoke AI consent at any time from Privacy & Data → Your consents .
• Export your data (GDPR Article 15) with the ai_generated flag intact.
• Request your AI audit log via our contact page .
• Ask us to erase your data, including cached AI responses, via Privacy & Data → Request data erasure on the web, or More → Delete my account in the iPhone app.