MEDILIT

Capture, redaction, note generation, the self auditing pass, and letter and document generation are all model output, delivered through a multi agent architecture the company describes as its core technology. No human scribe tier, no services layer and no host platform that the AI is a feature of.

The oversight structure is described, and one part is missing. The explanation MEDILIT publishes for patients states that after the consultation the doctor carefully reviews the draft, makes any necessary edits for accuracy and completeness, and approves the final version for the medical record. Alongside that, a self auditing layer checks the output for omissions and consistency before the clinician sees it.

The self audit is an internal consistency check run by the same system that produced the output, not independent verification, and it does not reduce the need for the clinician's review.

What is missing is what happens at the edges: no confidence signal, abstention behavior or flag is described for when the audit layer finds a problem or when the clinician should look harder, and nothing says how an error found after approval is corrected.

The approach is named, and version discipline is not. MEDILIT states that it uses its own AI models, developed for healthcare specialties and fine tuned on medical data and consultations, and describes a multi agent architecture with a self auditing layer intended to catch omissions and hold consistency across long consultations. A published AMD case study names the processor family the product runs on, and the subprocessor list includes Hugging Face.

No model card, base model, version, update practice or evaluation methodology is published. The accuracy figures, 98 percent plus precision on the home page and 95 percent from user surveys in the FAQ, carry no method, and the headline calls the product the most accurate AI medical scribe with nothing to support the comparison. The models are described as continuously learning from each interaction, which makes version and update disclosure more important rather than less. A self auditing claim with no measured catch rate is an architecture description, not a result.

Substantial partial disclosure with a short model chain. The company publishes a subprocessor list naming Microsoft Azure, Google Cloud, AWS, GitHub, Hugging Face and Stripe, and states that it uses its own AI models, fine tuned on medical data for healthcare specialties, rather than a third party language model. It also describes a multi agent architecture with a self auditing layer, and a published AMD case study names the processor family the product runs on.

The chain is bounded by design as well as disclosure. Identifiers such as names, dates of birth, numbers and email addresses are redacted on the clinician's device before transcripts leave it for processing, so downstream parties receive content with direct identifiers removed.

The roles are not assigned: the list does not say which cloud processes Australian clinical content, what Hugging Face is used for, or what retention and training terms bind each party. Redaction of identifiers spoken conversationally, a patient naming their street or employer mid sentence, is the hard case and no performance is published for it.

Vendor reported figures with no method behind them. The site claims 98 percent plus precision in note generation, its FAQ cites a 95 percent accuracy rate from user surveys, and it reports users saving 1.5 to 2 hours a day and up to 1.8 more patients an hour, with no denominator, method or reference standard for any of them. The company publishes no study, controlled evaluation or named customer.

The AMD case study is an externally published document, but it is a technology partner's marketing material rather than clinical evidence, describing time savings with no measurement, cohort or comparator.

A privacy design rather than only a policy, with the training side left open. Identifiers such as names, dates of birth, numbers and email addresses are redacted on the clinician's device and replaced with random values before transcripts leave it, so the generation layer does not receive them. Voice recordings are never stored; speech is converted to text in real time, transcriptions are deleted once the note is generated, and outputs are held only until they are transferred to the record system. Data is encrypted in transit, at rest and after processing, and processing stays on Australian servers.

The training element has no retention period behind it. The deletion commitments are scoped: the home page states that audio is processed momentarily and not stored, and that the doctor approved note becomes part of the official medical record under the practice's own retention rules. So audio and transcripts are covered and the note belongs to the practice. What is not covered is what the models learn. The same page describes the models as fine tuned on real medical data and consultations and continuously learning and improving from each interaction, and the subscriber agreement grants MEDILIT an irrevocable license to de identified data derived from customer health information, owned by MEDILIT and usable without limitation, with no retention window attached.

The question is narrower and harder than it first looks. It is not that the deletion claims and the learning claim cannot both be true; de identified derivatives can outlive the transcripts they came from. It is that everything the vendor commits to deleting has a clock, and the one category it keeps permanently has none.

No safety engineering is published beyond the self auditing layer: no guardrails, hallucination mitigation or handling of a safety event.

The vendor invokes the United States regime alongside Australian law. MEDILIT is Australian, operates under the Privacy Act 1988 and the Australian Privacy Principles, and states that processing stays within Australia, but its site also carries a HIPAA compliant badge and its privacy statement includes a HIPAA compliance addendum. The addendum names a designated HIPAA privacy and security officer, commits to administrative, physical and technical safeguards, regular risk assessments, breach notification within 60 days of discovery, annual staff training and six year retention of documentation including business associate agreements. The privacy statement describes the company as a business associate, and the subscriber agreement commits both parties to executing a business associate agreement.

The agreement itself is not published, and the subscriber agreement defines it by reference to the Australian Privacy Act rather than HIPAA. No United States offering is described, and processing is committed to Australian servers.

Recognized certifications are named in the vendor's own material, inconsistently. The compliance section of the site says ISO 27001 certified, the security summary on the same site says ISO 27001 compliant and SOC 2 certified, and the footer says SOC 2 compliant. The company also states that it conducts regular internal and external audits against the Australian Privacy Act, and offers a resources request route from its trust center section. No certificate, report type, date, auditor or scope is published.

The described controls are specific: voice recordings are never stored, speech is converted to text in real time, identifiers are redacted on the clinician's device before transcripts leave it, transcriptions are deleted once the note is generated, outputs are held only until transferred to the record system, and access is logged and audited.

A pipeline that claims never to store audio and to delete transcripts on note generation is exactly what an independent assessment exists to test, so the difference between certified and compliant matters here.

No device pathway attaches and none is claimed, and the vendor states its regulatory position precisely: MEDILIT says it is exempt from registration with the Therapeutic Goods Administration of Australia and that its platform and services are not considered a medical device.

Scope supports that. The product produces notes, letters, care plans and forms including Medicare and Department of Veterans' Affairs forms, and no diagnostic suggestion or order drafting is described. Care plans are marketed as meeting Medicare requirements, which is a payment documentation claim rather than a device question.

The vendor also publishes an explanation written for patients: what the tool is, that it converts the conversation to text without recording audio, that only the treating clinician and authorized staff can see the output, that access is logged, and that use is voluntary, declining does not affect care, and the patient can change their mind at any point. In Australia, privacy law, the regulator's guidance on digital scribes and medical defense organization positions all converge on informed patient consent, and addressing the patient directly is the harder half of that.

The gap to track is market expansion, since none of this transfers to a jurisdiction treating ambient documentation as a regulated device.

One real governance positive and one complete gap. The positive is structural: choosing to strip identifiers before the analysis layer is a data governance decision taken at design time rather than a control bolted on afterward, and it limits what the vendor itself can see.

The gap is the absence of any fairness statement, subgroup analysis, accent or dialect performance disclosure, or model limitations document, while the product markets itself on accuracy, on handling complex and lengthy consultations, and on supporting more than 120 languages.

Care plan and Medicare form generation is marketed on meeting Medicare requirements and on converting non billable work into billable items, which ties documentation to payment, though the site uses no coding intensity or reimbursement optimization language.

A review step is published and nothing stands behind the output. The explanation MEDILIT publishes for patients states that after the consultation the doctor carefully reviews the draft, makes any edits for accuracy and completeness, and approves the final version for the record. The architecture also includes a self auditing layer intended to catch omissions, though a self audit run by the system that produced the output is not independent verification.

The subscriber agreement warrants that the services will perform as described, with re-performance as the exclusive remedy, disclaims other warranties, caps liability at twelve months of fees and excludes gross negligence and fraud from the cap. Enterprise customers can cancel and receive a refund of the remaining balance. None of that attaches to the accuracy of a note.

The accuracy figures do not help a buyer. The site cites 98 percent plus precision in note generation, its FAQ cites a 95 percent accuracy rate from user surveys, and its headline calls it the most accurate scribe on the market, with no method, denominator or comparison behind any of them.

The integration claims contradict each other, and none names a connected system. MEDILIT is a browser based application. Its site says it works seamlessly with all practice management software used in Australia and New Zealand, that it is built to work alongside that software without requiring any integration, and elsewhere lists EMR integration with seamless transfer to the patient file, while its FAQ says it aims to partner with all Australian practice management providers and will update users as each integration is finalized. No system is named as connected, and no interface, standard or transfer mechanism is identified.

This matters beyond workflow because the retention design depends on it. Transcriptions and outputs are held only until successfully transferred to the record system, after which nothing is kept. Where no integration exists and a clinician copies text out of a browser by hand, it is not clear what confirms transfer or when deletion fires.

Residency is committed and the processing path is partly described, with isolation left open. MEDILIT is a browser based application with no local installation. It states that all processing happens on secure servers within Australia and that data never leaves Australia, that identifiers are redacted on the clinician's device before transcripts leave it, and that cloud storage is used only temporarily until documents are transferred. Its privacy statement adds that data is initially stored in Australia and that any transfer offshore would be handled under Australian privacy law. Its subprocessor list names Microsoft Azure, Google Cloud and AWS without assigning roles.

A published AMD case study documents the product running on AMD Ryzen Embedded 8000 Series processors, which does not match a browser only product.

Tenant isolation is not described, the privacy statement leaves room for offshore transfer that the marketing rules out, and the list does not say which cloud holds Australian clinical content.

A price is published with the route to start, without every tier priced. The Premium plan costs A$125 a month or 960 a year, excluding GST, and is designed for individual users. Standard and Premium both include the AI scribe plus letter and custom document generation, and Premium adds unlimited AI document generation and additional assistant tools. The trial grants full premium access with no payment details, and plans can be canceled at any time. An enterprise plan for practices with three or more doctors is annual, with discounts of up to 30 percent by group size and a refund of the remaining balance on cancellation.

The Standard plan's price is not stated, and enterprise pricing requires a quote.

Coverage is stated clearly and nothing validates it yet. The site lists more than twenty specialties by name, among them general practice, neurology, obstetrics and gynecology, pediatrics, psychiatry, emergency medicine, cardiology and oncology, and claims 40 more. For general practice it names specific Australian instruments: GP chronic condition management plans, mental health care plans, Department of Veterans' Affairs forms and Medicare forms, built to Medicare requirements and RACGP standards. The document range is wide, covering consultation notes, letters, referrals, patient instructions, reports, plans and custom documents, with the design aimed at long and complex consultations.

No validation sits behind any of it: no specialty specific accuracy, no evaluation in a named setting and no named customer, and the Australian and New Zealand general practice focus is the only depth the materials support.