AI Transparency Statement

How Praxi uses AI, what data is processed, where it runs, and where the human still decides.

House of Praxi Pty Ltd ABN 75 693 330 522 | ACN 693 330 522

Effective Date: 19 February 2026 Last Updated: 30 September 2026

Our Commitment to Transparency

At Praxi, we believe that transparency about how we use artificial intelligence is essential to building trust with the practitioners, practice staff, and clients who rely on our platform. This statement explains what our AI does, how it works, what safeguards are in place, and what it cannot do.

How Praxi Uses AI

Praxi is an AI-powered practice management platform designed for allied health practitioners in Australia. Our platform integrates six specialist AI agents that assist practitioners with administrative, clinical, research, and operational tasks. Every AI feature is designed to assist practitioners — never to replace their professional judgement.

What Our AI Agents Do

Clinical assistant (shown in the product as Praxi; the product source calls this function beck-clinical). What it does: drafts clinical notes and session summaries from what the practitioner said or wrote, drafts reports and letters for the practitioner to edit, and states an assessment's total score with the instrument's published score band and citation. What it does not do: it does not suggest a diagnosis or an ICD-10-AM code, write a treatment plan, rate or assess risk, write a safety plan, or recommend an intervention, client tool, assessment or care pathway. It does not sign clinical notes, and it does not write to the record until a practitioner approves the output. Practitioner approval does not decide whether the function is a medical device.

Jarvis (Operations). What it does: Assists with scheduling, billing queries, Medicare item lookups, and general platform help. What it does not do: Does not process payments autonomously or access external systems without practitioner authorisation.

Ada (Research). What it does: Summarises research evidence, looks up diagnostic criteria and MBS item numbers, creates psychoeducation materials. What it does not do: Does not provide medical advice, guarantee research accuracy, or replace professional literature review.

Warren (Analytics). What it does: Analyses practice performance metrics, identifies trends, assists with marketing content. What it does not do: Does not make business decisions or access data outside the practitioner's own organisation.

Brene (HR and Wellness). What it does: Monitors staff workload indicators, assists with leave management, supports onboarding. What it does not do: Does not make HR decisions, access information outside authorised scope, or replace professional HR advice.

Linus (Automation). What it does: Creates workflows, configures integrations, sets up automation rules. What it does not do: Does not modify system security settings or execute automations without practitioner approval.

The AI Receptionist (client-facing)

Separately from the practitioner-facing functions above, we are developing the AI Receptionist — a text-based reception feature that is coming soon and is not yet available to practices. It is intended for routine practice chat, SMS and email enquiries. The design safeguards described below apply to this feature in development:

• It always identifies itself as an AI assistant. It never claims to be human and does not use a personal name.
• It is enabled by the practice, never by default. Each practice chooses to turn it on and configures what it may handle.
• It does not give clinical advice. It handles reception matters (appointments, practice information, routine questions) and hands the conversation to practice staff when a message is outside that scope, suggests urgency or distress, or when the person asks for a human.
• Email replies are drafts. AI-prepared email responses are queued for practice staff to review before anything is sent.
• Every conversation is logged and visible to practice staff, with the same tenant isolation and audit trail as the rest of the platform.

Voice Transcription

With the client’s recorded consent, session audio can be transcribed to help draft clinical notes. Praxi checks that recording consent is on file before accepting any audio, and checks the separate AI-processing consent before any AI processing. Praxi does not keep the audio, transcripts are minimised after the related note is signed, and processing occurs within Australian AWS infrastructure (ap-southeast-2). The practitioner reviews and signs every note drafted from a transcript.

What AI Is and Is Not Used For

AI is used to:

• Draft clinical documentation for practitioner review
• Suggest scheduling arrangements for practitioner approval
• Summarise research and evidence for practitioner reference
• Generate administrative content (invoices, reminders, reports) for practitioner review
• Create workflow automations designed and approved by the practitioner

AI is never used to:

• Make clinical decisions on behalf of a practitioner
• Suggest a diagnosis, rate risk, or recommend a treatment or intervention
• Automatically execute any action affecting client care
• Replace the professional judgement of a registered health practitioner
• Access or process data without appropriate authorisation
• Communicate with clients unless the practice has explicitly enabled a client-facing feature (such as the AI Receptionist) — and never while presenting itself as human

Human Oversight and Practitioner Control

This product design supports (but does not replace) practitioners’ duties under Ahpra and the National Boards’ guidance *Meeting your professional obligations when using Artificial Intelligence in healthcare* (August 2024): accountability, understanding, transparency, informed consent, and ethical and legal issues.

Every AI Suggestion Requires Approval

This is the most important safeguard in our AI system. No AI-generated output is acted upon without the explicit approval of the practitioner. The process works as follows:

• The practitioner makes a request (e.g., "Draft a progress note for today's session")
• The AI generates a suggestion based on the available context
• The suggestion is placed in the practitioner's Outbox for review
• The practitioner reviews the suggestion and chooses to approve, edit, or reject it
• Only after the practitioner takes action is the suggestion executed

This applies to every practitioner-workflow AI feature, regardless of how routine or low-risk a suggestion may appear. The one deliberate, clearly-bounded exception is the client-facing AI Receptionist: once a practice has explicitly enabled it, it responds to routine reception messages (chat and SMS) directly, within the guardrails described above — it identifies itself as AI, gives no clinical advice, escalates to staff, prepares email replies as drafts only, and every conversation is logged for practice review.

Practitioner Responsibility

Practitioners who use Praxi's AI features retain full clinical and legal responsibility for:

• Reviewing all AI-generated content before approval
• Applying their professional judgement to every AI output
• Ensuring the accuracy and appropriateness of any AI content they approve
• Reporting any AI outputs they consider unsafe, inaccurate, or inappropriate

This aligns with AHPRA's Meeting your professional obligations when using AI in healthcare guidance (issued 22 August 2024 by AHPRA and the National Boards). See our AHPRA Practitioner Notice for further detail on the division of responsibility between Praxi and the practitioner.

Configurability

Practitioners and practice owners have control over AI features through a three-tier settings system:

• Practice Level: Practice owners can enable or disable specific AI features, set cost limits, and configure default behaviours for their organisation
• Practitioner Level: Individual practitioners can adjust AI preferences such as output format, tone, and notification settings
• Platform Level: Praxi maintains baseline safety settings (such as Australian-only clinical inference and redaction wherever information leaves that clinical boundary) that cannot be overridden

Data Privacy in AI Processing

How We Protect Your Data

Praxi takes extensive measures to protect personal and health information during AI processing:

• Clinical inference and boundary redaction: Drafting a note or letter needs the client's name and clinical content, so identifiers are not stripped before the model sees them. Instead we constrain where that processing may happen: clinical AI runs only on Anthropic's Claude models through AWS Bedrock in Sydney (ap-southeast-2), under terms that exclude your data from model training, inside the same tenant-isolated, encrypted, audit-logged boundary as the rest of your record. Personal information is redacted whenever it leaves that clinical boundary — for example analytics and log storage, feedback you send us, or a report to a recipient who is not entitled to the full record.
• Australian Data Residency: Clinical AI processing occurs through AWS Bedrock in the Sydney (ap-southeast-2) region. Your data remains within Australian infrastructure.
• No Model Training: Your data is never used to train AI models. We use Anthropic's Claude models through AWS Bedrock, which guarantees that customer data is not used for model training or improvement.
• Encryption: Data is encrypted in transit using TLS 1.2 or higher, with TLS 1.3 where supported, and sensitive stored fields use AES-256-GCM. AI interactions are stored in the same encrypted, tenant-isolated database as all other platform data.
• Tenant Isolation: Every organisation's data is isolated through PostgreSQL Row-Level Security (RLS). AI agents can only access data belonging to the practitioner's own organisation.
• Audit Trail: Every AI interaction is logged in an immutable audit trail retained for 7 years, recording what was requested, which agent and model processed it, and what action the practitioner took.

Third-Party AI Providers

For clinical and practitioner-facing AI, Praxi uses Anthropic's Claude AI models, accessed exclusively through AWS Bedrock. We have selected this arrangement because:

• AWS Bedrock provides enterprise-grade security and compliance (SOC 2 Type II, ISO 27001, IRAP)
• Data processed through Bedrock is not shared with Anthropic and is not used for model training
• Processing occurs within the AWS Sydney region, maintaining Australian data residency
• AWS Bedrock's data processing terms align with Australian privacy requirements

Support and sales phone line. Our customer support and sales phone line (+61 480 890 089) uses ElevenLabs Conversational AI (United States) for voice and its own language model, separate from the clinical AI described above. This line is for software support and sales. Please do not share patient or clinical information; details supplied during a call may be processed by ElevenLabs. The client-facing AI Receptionist, which is coming soon, does not use ElevenLabs: it is text-based and runs on the same Anthropic Claude via AWS Bedrock (Sydney) arrangement described above, with the same no-model-training and data-residency protections.

For the full list of third parties we work with, see our Sub-processors page.

Automated decision-making transparency (POLA)

The Privacy and Other Legislation Amendment Act 2024 (Cth) introduces automated decision-making transparency obligations for APP entities. Those obligations are due to take effect on 10 December 2026.

In anticipation of that commencement, Praxi already discloses the kinds of processes that are substantially automated in our Privacy Policy (including the decisions Praxi's software makes without a person deciding each case, such as sign-in lockouts, security restrictions pending review and subscription access). Praxi does not use software to decide anyone's assessment, diagnosis, treatment or eligibility for care. Where AI flags a concern or generates a recommendation, a practitioner or administrator reviews the outcome before it is acted on.

You may request human review of any automated output that affects you by contacting our Privacy Officer at hey@praxi.au. Clients of Praxi practices can also use the “Your data and AI” section in the client portal account settings.

Related OAIC guidance on deploying commercial AI products (due diligence, human oversight, APP 3/6/10/11) is at the OAIC commercial AI products page.

Limitations and Disclaimers

AI Is Not Perfect

We are transparent about the limitations of AI technology:

• AI can be wrong. AI models may generate inaccurate, incomplete, or inappropriate outputs. This is why every output requires practitioner review before approval.
• AI is not a substitute for clinical judgement. AI suggestions are based on patterns in training data and the context provided. They do not account for all factors a trained practitioner considers in clinical decision-making.
• AI may reflect biases. Despite our efforts to mitigate bias, AI models may reflect biases present in their training data. Practitioners should be aware of this and apply their professional judgement accordingly.
• AI does not understand context the way humans do. AI processes text statistically. It does not have genuine understanding of clinical situations, therapeutic relationships, or the nuances of individual client presentations.
• AI knowledge has a cut-off date. AI models are trained on data up to a specific date and may not reflect the most recent research, guidelines, or regulatory changes.

Regulatory Status

Praxi's AI features are designed as assistive tools for practice management and clinical documentation. Regulatory classification depends on the intended purpose and functionality of each feature and the applicable TGA criteria. Practitioner oversight alone does not establish an exclusion or exemption; this statement does not claim a blanket medical-device classification for all features. Praxi's AI does not suggest a diagnosis, a treatment plan, a risk rating, a formulation or a tailored intervention for a named client. It drafts documentation from what the practitioner said or wrote, and the practitioner reviews and signs it. Practitioner review does not, by itself, take a function outside the definition of a medical device. We reassess a function's intended purpose whenever it changes, and against the clinical decision support exemption as reworded from 1 November 2026. This statement does not say that inclusion in the Australian Register of Therapeutic Goods is required.

Reporting Concerns

If you have concerns about any AI output or behaviour within Praxi, please report it to us:

• In the platform: Use the feedback button available on any AI output to flag the issue
• By email: hey@praxi.au
• Privacy concerns: Contact our Privacy Officer at hey@praxi.au
• Security concerns: security@praxi.au — coordinated security disclosure is welcome; see our Breach Notification Commitment and Security Controls pages

We take all reports seriously and investigate every concern raised about AI safety, accuracy, or appropriateness.

Updates to This Statement

This statement will be updated as our AI features evolve and as the regulatory landscape develops. Material changes will be communicated to all platform users. The current version is always available at praxi.au/legal/ai-transparency. We review this statement at least annually, and sooner whenever a new AI feature materially changes how AI is used on the platform. Review scheduling is tracked in our internal compliance register.

Contact

House of Praxi Pty Ltd ABN 75 693 330 522 | ACN 693 330 522 493 Swan Street, Richmond VIC 3121 Email: hey@praxi.au Website: praxi.au