Conveyor AI by Mobius MD

The ambient generation is genuine model work, but the company's patented asset and the reason clinicians choose it are not AI. Mobius MD's differentiator is the Conveyor pairing technology, a connectivity invention from 2017 that lets a phone type into any computer without software installation, and the AI scribe was layered onto it afterwards.

That is the moat is not the model pattern this index has applied to Haystack and Armis: strip out the ambient AI and a valuable universal dictation business remains, which is not true of the pure plays here.

Conventional draft and review with a structural gate similar to River Records: because output is streamed to wherever the clinician has placed the cursor rather than written into the record through an integration, a clinician is necessarily present and in the document at the moment text arrives. Nothing files itself.

Held at B because no acceptance rate, edit burden figure or confidence threshold is published, and because generating 80 percent of a note quickly says nothing about how much of the remaining review burden is correction.

The headline figure measures completeness rather than correctness: more than 80 percent of the note generated within 60 seconds says how much text arrives and how fast, not how much of it is right. Specialty specific models are referenced without being described, and no accuracy figure, model card, named model or evaluation methodology was located. Mild credit for a public blog that discusses how ambient scribes work and where they fall short, though that is category education rather than disclosure about this product.

Nothing identifies any party in the chain: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located in two passes. Specialty specific models are referenced without being described, so a buyer cannot establish whether those are distinct trained models, configuration of one model, or prompt templates, and the difference matters for who is in the chain.

The mobile platform restriction is the only architectural commitment published, and it bounds the client device rather than the processing path behind it. Ask whether any third party model provider is invoked, for a sub processor list, and for what the specialty specific models actually are.

Named clinician testimonials across orthopaedic surgery, internal medicine, family practice, emergency and telemedicine, plus ten verified clients recorded on a third party marketplace listing. No study, controlled comparison, accuracy benchmark or independent evaluation located. Graded C on the standing precedent that testimony and client counts are not evidence of benefit.

Claims to exceed the standards specified by HIPAA and HITECH, which is an assertion rather than a control. What lends it some weight is a concrete and costly design decision: the product requires an iOS device and supports no Android path at all, which the company attributes directly to its security requirements. Giving up an entire mobile platform is an unusual thing to do for security theatre, so the constraint reads as real. No retention schedule, de identification practice or training use statement was located.

States compliance with HIPAA and HITECH, the latter being a reference few vendors in this category make explicitly. Business associate agreement terms are not published for inspection before contracting.

No SOC 2 report of either type, no HITRUST certification, no ISO 27001 and no trust centre were located in a second pass.

The product constraint the earlier assessment noticed is now explained by the vendor, and it is worth examining rather than accepting. The application runs only on Apple mobile devices, and the vendor states this is due to stringent security requirements. That is a coherent argument on its face, since a single tightly controlled mobile platform offers consistent hardware backed key storage and no sideloading. It is also a design decision presented as a security conclusion, with nothing published to show the analysis behind it or what was rejected. A restriction is not an assurance, and a buyer should not read platform exclusivity as evidence of a posture.

The compliance formulation used alongside it belongs with the non attestations this index tracks. The vendor states the product exceeds the standards specified by the health privacy and health technology statutes. Those are laws rather than benchmarks, exceeding them is not a measurable claim, and there is no certifying body for either. It sits with military grade, enterprise grade and bank level in the same family.

What is concrete is worth crediting. The hardware dongles are described as bound to a single user account with an encrypted connection to that user's device, and the housing carries an antimicrobial additive, which addresses a real infection control problem with shared workstation microphones that no other vendor in this lane has considered.

The organisation has operated for over a decade with a stated United States based team, which raises the expectation that documentation exists.

Ask which report is held, of which type and period.

No clearance claimed and none required for ambient documentation or dictation. No United States device pathway attaches.

This record is one of the narrowest in the category, and that is worth stating as a positive rather than recording as an absence. The scope expansion pattern documented across most of this lane, meaning coding engines, risk adjustment scoring, decision support, order generation, patient facing products and revenue cycle modules, is absent here. No coding engine, no clinical insight layer, no patient facing component and no billing function were located. The product captures speech and produces text.

The delivery method reinforces how narrow that is. Output arrives wherever the clinician places their cursor rather than into a defined clinical field, so the product has no view of the chart, no access to patient context, and no capacity to act on the record. It is a text input method with a clinical vocabulary. That keeps the regulatory position simpler than almost any peer's, and following the treatment this index gave Nabla and OrbDoc, the narrowness is credited rather than read as immaturity.

One consequence runs the other way and belongs here. Because output is not scoped to a clinical destination, nothing in the product constrains where dictated clinical content ends up. That is a matter of practice governance rather than device regulation, but a buyer should establish what their own policy says about dictating patient information into applications outside the record system, since the tool makes it equally easy.

Markets appear to be United States only, with a stated domestic support and engineering base.

A second search confirms the absence. No fairness statement, no subgroup analysis, no accuracy breakdown by accent, dialect or speaker population, no evaluation methodology, and no language coverage claim were located.

The earlier assessment's point stands and is the right frame: this product's lineage is medical dictation, a field where accent handling has been a documented weakness of speech recognition for decades. A vendor building on that lineage is working on a problem the field has known about for a very long time, and is silent about where it stands on it.

The second pass adds a mechanism that deserves its own attention, because it is presented as a benefit and carries a governance consequence. The system adapts to the individual user. Custom vocabulary can be added, a specialty dictionary selected, and the vendor states that the more it is used the better it gets, with the trained voice profile travelling with the clinician between employers and workstations.

Personal adaptation is genuinely useful and it also obscures the thing this axis measures. An adaptive system improves toward each user, which means published accuracy reflects some point along that curve rather than a baseline, and it means the burden of adaptation falls hardest on the users the base model handles worst. A clinician whose speech the underlying model was trained to recognise starts near their ceiling. A clinician with a less well represented accent starts lower and spends longer correcting output before the system catches up, and that unpaid correction work is invisible in any headline figure.

So the disclosure that would matter here is not a single accuracy number but a starting position and a curve. Ask what performance looks like on day one, for whom, and how long adaptation takes.

Two passes located no accuracy or error figure, no published limitations and no warranty, indemnity or remediation commitment. The headline figure measures the wrong thing for this axis: more than 80 percent of the note generated within 60 seconds describes how much text arrives and how quickly, not how much of it is correct, and a buyer reading it as a performance claim would be reading it as something it does not say.

A pattern this index has recorded elsewhere appears again here in a milder form. The company publishes a blog discussing how ambient scribes work and where they fall short, which is useful category education, and publishes nothing about where its own product falls short. Writing about the failure modes of a technology while declining to characterise your own instance of it leaves a reader better informed about the category and no better informed about the purchase.

One design decision does read as substantive rather than promotional and is worth crediting on the record: the product requires an Apple mobile device and supports no alternative path at all, attributed by the company to its security requirements. Giving up an entire mobile platform is an expensive thing to do for appearances, so the constraint is probably real, though it speaks to data protection rather than to output. Ask for an accuracy figure with a definition, and for what the vendor commits to when a note is wrong.

Universal by mechanism rather than by integration, which is a real distinction a buyer should understand. Patented pairing lets an iOS device stream text into any application on any Mac or Windows computer with nothing installed, so it works with every EMR including Epic, Cerner and MEDITECH without an integration project, and a clinician moving between hospitals, systems or employers keeps their profile, templates and preferences.

That portability is a genuine anti lock in argument almost nobody else in this category makes. The limits are equally real: text lands at the cursor, so there is no structured field population, no chart context flowing in and no write back, and integrates in third party listings should be read as works alongside rather than connects to.

No hosting region, residency option or subprocessor list was located, and nothing establishes where transcription and generation run or whether a third party model service is involved.

The architecture is unlike anything else in this lane and needs describing, because it changes what the question even is. Capture happens on an iOS device. The text is delivered to a computer either through a thumb sized hardware dongle or a scanned pairing code, and appears wherever the cursor is placed. No software is installed on the computer and no integration with the record system exists. From the record system's perspective, a user is typing.

Two consequences follow that a buyer will not find on a diagram.

The first is scope. Because delivery is to the cursor rather than to a chart, the tool works in any application. The vendor markets this explicitly, naming documents, email and general web use alongside the record system. That is genuine flexibility and it also means clinical content can be dictated into destinations with no protection at all, at the user's discretion and outside any administrator's control.

The second is attribution. Text arriving as keystrokes carries no marker distinguishing it from typing, so the record contains no indication that any part of a note was machine generated. That is not a defect in itself, since the clinician is genuinely the author of what they send, but an organisation that expects to identify AI assisted documentation in its own records will not be able to.

One control is specific and creditable: the hardware dongles are described as bound to a single user account and establishing an encrypted connection to that user's device.

Ask where transcription and generation run, and what an administrator can see or restrict.

More transparent than most and not fully published. A seven day free trial with full feature access is offered, the company states there are no costs or surprise fees and that features other platforms charge extra for are included, and roughly 149 US dollars a month is cited for unlimited premium dictation.

Held at B because that figure appears in the vendor's own blog rather than a rate card, and pricing for the AI scribe tier and enterprise options is described as layered on top without being stated, so a buyer evaluating the ambient product specifically still cannot see the number.

Broad by user base rather than by enumeration, with documented use across orthopaedic surgery, internal medicine, family practice, emergency department, inpatient and telemedicine, and a dictation heritage that covers operative notes as well as office visits. Customisable templates and phrase libraries carry per clinician. Two bounds: an iOS device is mandatory, and no language coverage beyond English was located.