Trust centre

DEMONSTRATION

This is a demonstration build on made-up data.No clinic uses it for patient care yet. Everything below describes how it is built today; the “Not ready yet” section lists what must happen before real patients.

The AI line

AI in Iris does paperwork. It does not make clinical decisions, and a clinician signs every AI draft before it goes into a record.

What the AI does

Writes up what was said in a consultation, when the patient has agreed — and nothing that wasn’t said.
Lays out a doctor’s own rough notes, and drafts letters from the signed note.
Suggests Medicare items the signed note already supports, quoting the words that support them.
Drafts reminders and answers practice-management questions.

What it never does

Diagnose, interpret test results, choose treatment or doses, or decide how urgent a patient is.
Give clinical advice to patients.
Write that a patient consented — consent is recorded by the clinician. If the AI tries, the sentence is removed and logged.
Save a draft to a record without a clinician signing it — and the server will not accept a signature until every flagged item has been ticked off.

Why the line is here. Australian law treats software that recommends a diagnosis or treatment as a medical device. There is an exemption for clinical decision support shown to health professionals — but the TGA says an AI-generated recommendation cannot use it, and a clinician approving the output does not change that. So in September 2026 we removed three AI features that crossed the line (AI urgency triage, AI explanation of results, AI care-pathway answers) and replaced two of them with rule-based checks that show the clinician exactly what they checked, what they found and which guideline they come from.

Every AI draft is checked before a doctor signs it. Iris matches each sentence to what was actually said, looks for medicines, doses and safety-net advice that were added or left out, and makes the doctor tick off each flagged item before the note can be signed. Leaving things out is the most common scribe error, so we look for it specifically.

Feature register

Every feature, its regulatory status and why. Kept in the product and checked on every release.

Feature	Status	Conditions that keep it there
Loading…

Automated decisions

From 10 December 2026 the Privacy Act requires practices to say when computer programs make, or substantially help make, decisions that significantly affect people. This is Iris’s part of that statement.

Loading…

AI function	Who uses it	Risk	Safeguards

Security

Sign-in is built to the Essential Eight Maturity Level 2 requirement for phishing-resistant multi-factor authentication (Australian Signals Directorate) and to the Australian Digital Health Agency’s session rules for clinical systems. Backups, patching and central logging are not yet independently assessed — see “Not ready yet”.

Signing in

Passkeys required for every staff member — phishing-resistant multi-factor authentication (Touch ID, Face ID, Windows Hello or a security key). On by default; if a practice switches it off, its compliance page shows a BLOCK. Single-use recovery codes are the fallback: they are not phishing-resistant, so every use alerts the practice manager and wrong codes lock the account.
Signed out after 15 minutes without activity, and after 12 hours regardless. Sign-out ends the session on the server.
Sign-in locks after repeated failed attempts. Every sign-in, failure and lockout is recorded.
Passwords stored as salted PBKDF2 hashes. (The demonstration still accepts a few plain-text demo passwords; the manager’s compliance page flags each one until it is replaced.)

Inside the system

Role-based write control on the server: only a GP can change the clinical record; reception cannot change clinical notes, whatever the screen shows. (In the demonstration the practice manager may also sign “on behalf of” a GP so the flow can be shown; each such entry is labelled and logged, and this is switched off for real use.)
Tamper-evident audit trail: every entry carries a keyed fingerprint of itself and the one before it, so an entry edited, deleted or inserted in the store shows up. Sign-in noise is kept in a separate log so it cannot push record-access entries out. Not yet anchored outside the system.
Strict security headers on every page; cross-site requests refused; no third-party trackers, fonts or analytics.
AI keys held on the server only; the browser never sees them.

Privacy and your rights

Under the Privacy Act and state health records laws (such as Victoria’s Health Records Act 2001), you can ask your practice for these. Iris is built so the practice can do them quickly.

See your record — the practice can export your complete record, including any AI drafts and the checks the doctor saw.
See who has looked at it — each time staff open your record in Iris’s screens, and every change, is logged with the person and time.
Correct it — the server refuses any change to a signed note; corrections are added as dated addenda.
Say no to the AI scribe — it only starts after you agree, and saying no changes nothing else about your care.

How long records are kept — at least 7 years after your last visit, or until age 25 for information collected when you were a child, whichever is later.
Recordings — consultation transcripts are not kept by Iris. The planned front-desk robot will record nothing by default.
Complaints — talk to your practice first. You can also contact the Office of the Australian Information Commissioner (oaic.gov.au) or your state health complaints body (in Victoria, the Health Complaints Commissioner, hcc.vic.gov.au).

Rural practices and outages

Connections drop in the bush. Care shouldn’t.

Offline-ready computers keep showing the day’s record through an outage, keep changes and send them when the connection returns — and never overwrite someone else’s change without asking.
Printable downtime pack each morning: today’s patients, allergies, medicines and results waiting for a decision.
One-page emergency transfer summary for retrieval teams.
Telehealth rules checked at booking (the 12-month face-to-face rule and its exemptions), practice-set appointment rules, and the interpreter line shown to the doctor when a patient asks for one.

Standards we build to

Alignment is not certification. Where we say “aligned in design”, no independent body has assessed it yet.

Framework	What it covers	Status
TGA medical device regulation (CDSS exemption, digital scribe guidance)	Where software becomes a medical device	Boundary enforced in code
Privacy Act 1988 (APPs), incl. automated-decision transparency from 10 Dec 2026	Collection, use, security, access, correction, overseas disclosure	Aligned in design · residency gap
Health Records Act 2001 (Vic) and NSW HRIP Act 2002	Health privacy principles, retention, transfer, practice closure	Aligned in design
ASD Essential Eight (Nov 2023 model)	Baseline cyber controls	MFA at ML2 · backups, patching, logging to do · device controls are the practice’s
ASD Information Security Manual; Secure by Design	Security controls and secure development	Partly
ADHA My Health Record Security Conformance Profile	Clinical systems connecting to My Health Record	Sign-in controls met · not connected
RACGP Standards for general practices (5th ed., C6.4; 6th ed. AI criterion)	Information security and AI use in practices	Supports the practice’s evidence
National AI Centre — Guidance for AI Adoption (six essential practices)	Accountability, impact, risk, transparency, testing, human control	AI register, kill switches, evaluation set
AS ISO/IEC 42001:2023; ISO/IEC 42005:2025	AI management system; AI impact assessment	Aligned in design · not certified
ISO/IEC 27001:2022; ISO 27799	Information security management	Planned
ISO 14971; IEC 62304; IEC 82304-1; IEC 81001-5-1; ISO 13485	Medical-device risk, software lifecycle, health software safety and security, quality system	Planned — required for any regulated feature
ACSQHC AI Clinical Use Guide; AHPRA guidance on AI	Safe clinical use of AI	Consent, review, reporting built in

Not ready yet

What must happen before any real patient information is entered. We would rather tell you than have you find out.

Critical Data location. The demonstration stores its (made-up) data on a global cloud key-value store that cannot be restricted to Australia. Real patient data needs an Australian-region database first.
Required Immutable daily backups with tested restores, central security logging with alerts, independent penetration test, provider contracts confirming no training on and no retention of data by AI providers, and a privacy impact assessment.
Required Written TGA advice on the booking-page urgent-symptom message, and TGA notification within 20 working days of first supplying the rule-based clinical checks.
Required A licensed Australian medicines database to replace the demonstration prescribing rules; the practice’s clinical governance to approve each care-pathway rule.
Required Live audio for the scribe stays off until there is an Australian speech-to-text service, a consent process and a privacy impact assessment.

Report a problem

Clinicians can report a problem with any AI draft or clinical check from the screen itself; the practice manager sees it with any regulatory reporting deadline.

Found a security issue? Please tell the practice manager, who will pass it to the Iris team the same day. Please don’t access data that isn’t yours or disrupt the service. A dedicated security address will be published in /.well-known/security.txt when the Iris company is registered.

EMP-063 · Iris trust centre · build 14 · · Practice home