Augnito

States plainly that it owns and trains its own AI models rather than orchestrating someone else's, and the evidence supports it: separate speech models tuned per medical specialty is a first party training decision, not something achievable by prompting a general purpose API. The parent's transcription heritage is a services business, but the product itself is model work.

No review gate between generation and transfer is described, and no acceptance rate, edit burden figure or confidence threshold is published.

This is more than a routine gap. The product does not stop at a note. It generates and sends clinical letters, captures clinical coding, and performs live translation during the consultation. Each of those carries consequences outside the chart, and each warrants a different answer about what the clinician confirms and when.

Outbound letters are the sharpest. A letter that has been sent has reached another clinician or a patient and cannot be recalled by editing the record later. Nothing published says whether the clinician approves each letter before dispatch, what the default is, or whether sending can be automated.

Translation is the second. Where the system renders speech across languages inside a live exchange, the clinician cannot verify the rendering unless they speak both languages, so the ordinary review gate does not function. No statement says the product cannot act as an interpreter or that a native speaker must corroborate.

The vendor markets quality assurance as a feature without describing what performs it or what it checks. If a quality assurance step exists it may already be a review gate, and if it is automated it is not oversight.

Useful architectural disclosure undermined by unfalsifiable accuracy language. On the positive side, Augnito describes a distinctive design: it owns and trains its models, and ships multiple speech models tuned to specific specialties rather than one general model behind specialty templates, which is a structural choice a buyer can interrogate. Against that, the accuracy claims are weak.

Spectra is marketed at 99.9 percent accurate speech recognition with no methodology, reference standard or denominator, and Omni's materials claim transcription with absolute accuracy, which is not a property any speech system possesses and cannot be tested. For a product sold on measurable accuracy, no measurement is published.

The corpus provenance is unusual, and it bears on the supply chain because the question of what fed the model is inseparable from who else has held the content. The company owns and trains its own models and ships several speech models tuned to particular specialties rather than one general model behind templates, which is a short chain by design and one a buyer can interrogate.

Where the training material came from is the substantive disclosure. The parent has supplied clinical transcription as a service since 2001 with delivery teams in England and India, and an account of the product's development describes a large existing database of medical voice recordings from doctors of many nationalities working in that health service as the basis of the speech engine's accent handling, which is the product's principal claim. So a corpus accumulated under service contracts appears to underpin a commercial speech product. That may be entirely proper, the contracts may permit it and the material may have been de-identified, and the speakers are clinicians rather than patients, but it is a change of purpose from the one the recordings were collected for, and dictated clinical content concerns patients even when the voice is a doctor's.

The parent's privacy policy at scribetech.co.uk/privacy-policy does name the parties it contracts with for service delivery, listing Scribetech India and Augnito India Private Ltd, and addresses transfers of personal data outside the United Kingdom under the applicable data protection regime. That is the corporate side of the chain rather than the technical one, and the document is scoped as a general privacy policy rather than a subprocessor list for the clinical product, so it does not establish which entity holds clinical content at which stage.

Still unenumerated: no hosting or cloud arrangement is published, no model or model family is named beyond the statement that the models are the company's own, and no subprocessor list covering the clinical service is published.

No peer reviewed study, controlled evaluation or independent assessment is published. NHS trust and private provider adoption is asserted across UK and global deployments, and the parent has a long NHS transcription history, but deployment history is not evidence of the AI product's benefit. Community based nursing deployment is described in case study form rather than measured.

No statement on audio or transcript retention, deidentification, or whether customer data trains models is published. That matters more here because the vendor trains its own models and competes on accuracy gained from training data. Where that data came from is the most consequential fact about this vendor's data handling.

The parent company has supplied clinical transcription to the health service and private providers since 2001, with delivery teams in England and India. An account of the product's development states plainly that one of the company's advantages was a large existing database of medical voice recordings, representing doctors from all over the world with varying accents who worked in that health service. That accumulated corpus is described as the basis of the speech engine's accent handling, and accent handling is the product's principal claim.

So recordings gathered over two decades under contracts to provide transcription as a service appear to underpin a commercial speech product. That may be entirely proper. The contracts may permit it, the material may have been deidentified, and the speakers are clinicians rather than patients. It is still a change of purpose from the one the recordings were collected for, and dictated clinical content is about patients even when the voice is a doctor's.

Nothing published says on what basis that corpus was used for model development, what deidentification was applied, whether the originating organizations agreed, or whether a current customer's dictation joins it.

No business associate agreement posture is published, and deployment is centered on the United Kingdom health service and international markets under the British data protection regime rather than the United States rule.

That scoping would make the United States rule largely irrelevant were the vendor's reach narrower. It is not. The company markets to global hospitals, launched first in India, and sells its speech engine to other software vendors who may deploy it anywhere including the United States. The vendor's home market materials give a United States customer, or a customer of a third party product built on this engine, no answer.

Two gaps follow.

First, nothing says whether an agreement is available at all. A vendor operating principally under a different regime may have no mechanism to offer one, and using the tool on protected health information without it is itself a violation however secure the product is. The same applies to any software vendor embedding this engine: their own agreement with a United States customer does not automatically extend to their supplier.

Second, and specific to this vendor, nothing says where processing and any human review occur. The company operates delivery teams in two countries and its origins are in human transcription, so the location of people, not only servers, is a live question. Under the United States rule an offshore human reviewing protected health information is a permitted arrangement only if it is papered, and no such paper is published.

No named or dated attestation and no trust center is published.

The expectation here is higher, and the reason is the customer base rather than the company's size. Supplying the United Kingdom health service brings a defined set of assurance requirements, principally the national data security and protection toolkit, which suppliers handling patient data are expected to complete annually and which produces a published, dated status. Beyond it, ISO 27001 and the national cyber essentials scheme are the usual accompaniments.

So the relevant evidence for this vendor is standardized, annual and published by design. A supplier with two decades of health service work will hold a toolkit status. That none of it appears on the product's own materials is a publication gap rather than, most likely, a compliance one, and it is straightforwardly fixable.

Two features of the business raise what an examination would need to cover. The company has historically delivered services through teams in two countries, so personnel security, access control and offshore working arrangements are substantive rather than boilerplate. And it sells its speech engine as an interface and development kit for third party clinical systems, so other vendors inherit this posture into their own products and pass it downstream.

Augnito Omni is published as an MHRA Class I registered medical device on the parent company's estate at scribetech.co.uk, alongside the statement that it is UK data trained and GDPR compliant. The document sits on scribetech.co.uk rather than augnito.ai.

A named registration with a named regulator is a stated regulatory position rather than silence.

Two problems remain, and the second is the vendor's own doing. The registration is asserted without a registration number, without the stated intended purpose, and without the scope or indication it covers, so it is not verifiable from what is published. And the same estate describes the product as delivering NHS approved suggestive feedback, with the health service named as approver. Approval or adoption by a health system is a procurement outcome, not a regulatory classification. No health service approves a medical device; a regulator registers one. Placing an institution's name in that position invites a reader to hear regulatory sanction where a purchasing decision is being described, alongside a real registration that makes a different kind of claim.

The intended purpose question remains live rather than settled by the class. Alongside notes, the product captures clinical coding in ICD-10 and SNOMED, generates and sends clinical letters, delivers documentation feedback, and performs live translation inside a clinical exchange. Nothing published says whether that set sits within the registered intended purpose.

The vendor engages accent variation specifically, and it asserts rather than demonstrates. Augnito does not gesture at multilingual support; it names the speaker populations it claims to handle, including Indian English, African English, Welsh, Scottish and Irish, alongside capture in more than 37 languages and per specialty speech models rather than one model behind specialty templates. Naming the populations instead of hiding behind a language count is meaningfully better than a bare count, and speech models do degrade on accents underrepresented in training data.

The problem is that the marketing converts a measurable property into an unfalsifiable one. The product is framed as accent agnostic and, in the parent's own material, as transcribing with absolute accuracy, which is not a property any speech system possesses. A vendor selling on accent performance is the one vendor positioned to publish word error rates broken down by accent, and none is published.

The accuracy claims across the company's products are also inconsistent, which sharpens the point rather than softening it: 99.9 percent appears on one product's materials and 99.3 percent on another's, neither with a methodology, reference standard or denominator, and neither broken down by the speaker populations the product is sold on.

The named accents and the absence of per accent measurement are both checkable.

No terms of service, warranty or indemnity language, remediation commitment or falsifiable performance measurement is published.

What is published instead is a set of accuracy claims whose weakness is of a particular kind that matters here. One product is marketed at 99.9 percent accurate speech recognition with no methodology, reference standard or denominator. Another is described as transcribing with absolute accuracy, which is not a property any speech system possesses and which no test could confirm.

A claim that cannot be true is worse than silence for a buyer trying to establish reliance, because it forecloses the conversation rather than opening it. And a vendor competing explicitly on measurable accuracy is the one vendor that should be publishing measurements.

The jurisdictional route, a correction right under the law where data is processed, applies only partially. Operations span England and India and the product sells across several markets, so whether the recorded patient holds an enforceable correction right depends on where a given deployment processes data rather than on anything inherent to the vendor.

Integration is offered as a capability rather than a fixed list, through an API and SDK designed to add speech and voice command into clinical software, EPR, PACS and RIS systems, alongside direct note and letter transfer into the electronic patient record. Radiology reporting systems are covered as well as record systems.

No named production integrations are published, so breadth is architectural rather than demonstrated. An API that could connect to a class of systems and a list of systems it is connected to today are different claims, and only the first is evidenced.

No hosting region, residency option or subprocessor detail is published. Delivery is cloud across desktop, browser and mobile with no voice profile training required.

Processing location is a specific question here rather than a generic one. The company operates delivery teams in England and India, and has done so throughout the two decades it has supplied clinical transcription. That is a stated operational fact rather than an inference.

For a health service buyer this is the first question their governance process will ask, and it has two parts that should not be collapsed. Where does automated processing run, meaning transcription and note generation. And separately, does any human ever see the content, and where are they. The second matters because this company's origins are in human transcription delivered from two countries, so a hybrid or fallback path is plausible in a way it would not be for a vendor with no such history. If any content reaches a person for quality assurance, correction or escalation, that is a transfer regardless of where the servers sit.

The vendor markets quality assurance as a feature of the ambient product without describing who or what performs it. That phrase should be resolved directly.

The interface product raises a third question. The company sells its speech capability as an interface and development kit for other software vendors to embed, so a clinician may be using this engine inside a third party clinical system, with residency governed by an agreement they never saw.

Publishes the pricing mechanism but not the price. Materials describe a simple recurring monthly fee with no capital investment required, which gives the shape of the commitment, but no rate, tier or per clinician figure is published.

Broad coverage, enumerated rather than implied. Spectra spans more than 55 medical specialties, Omni captures in clinic, over video call and via telemedicine, and coverage explicitly includes general practice, nursing including community based nurses, psychiatry, surgery and cardiology, with a speech model tuned per specialty rather than one model reused. Capture runs to more than 37 languages with live translation. The community nursing use case, where the clinician is rarely near a desktop, is addressed explicitly.