Ember Copilot

Models perform the review, the root cause analysis, and the appeal drafting. Coding checks against CPT, ICD-10, HCPCS, modifiers, and NCCI edits, plus payer policy and contract terms, are model outputs, and ambient scribing runs on the same platform.

The design answer to the problem this index has flagged repeatedly in revenue cycle AI. Where Waystar autonomously generates appeal letters with no published review step, Ember returns suggested corrections that carry the underlying rule and cite its source, from CMS, NCCI, or payer policy. A coder can verify why a change was proposed rather than accepting an unexplained flag, which makes the output auditable and the reviewer accountable. Citing the rule is the correct pattern for coding assistance.

No foundation model is named, no architecture is described and no evaluation methodology is published.

Real credit is due on explainability, and it is graded fully on the autonomy axis: outputs carry the underlying rule and cite its source, so a coder can see why a change was proposed rather than accepting an unexplained flag. One co founder's background is in explainable AI research, and the product design reflects it.

The grade stays at C because explaining a rule is not the same as disclosing a model, and the headline number is unsupported. A claim of 98 percent coding accuracy carries no audit standard, no denominator, no specialty breakdown and no independent validation, and coding accuracy is meaningless without stating what it was measured against and by whom. For a product performing autonomous coding, the two numbers worth publishing are accuracy against a credentialed coder audit and the rate at which autonomously coded claims are later denied or reversed.

The vendor states plainly what most in this index leave unsaid, and the plainness is worth crediting even though what it discloses is not the commitment a buyer would prefer. It generates aggregated or de identified analytics to improve model accuracy and platform reliability, subject to applicable law and the customer's agreement.

That is an affirmative statement that customer derived data improves the models, and the qualifier makes it negotiable rather than assumed, which is the practical difference between a disclosure and a fait accompli: a buyer who reads it knows there is something to ask to switch off. Two gaps hold this in the middle band.

No de identification standard is named, so what survives that process is defined by the company rather than by a stated method, and this index has recorded elsewhere that an undefined de identification claim leaves an entire commitment resting on one word. And nothing is enumerated: no model or model family, no foundation model provider, no hosting arrangement and no sub processor list was located, alongside no retention period for any data type the platform holds. Ask which de identification method is used, what the agreement permits you to decline, and for a sub processor list.

Specific outcome claims are published, 55 to 57 percent fewer denials and 98 percent coding accuracy, but they are vendor reported without methodology, baseline, or customer references, and the company is early with a $4.3 million seed in November 2025. Coding accuracy in particular is meaningless without stating the audit standard it was measured against.

The company discloses what most vendors in this index leave unstated: it generates aggregated or de identified analytics to improve model accuracy and platform reliability, subject to applicable law and the customer's agreement. That is a plain statement that customer derived data improves the models, and the qualifier makes it negotiable rather than assumed.

Alongside it: encryption in transit and at rest, customisable role based access and compliance settings, and stated audit trails. The disclosure and those controls together are why this sits at B rather than C.

Two gaps hold it there. No de identification standard is named, so what survives that process is defined by the company rather than by a stated method. And no retention period is published for any of the three data types the platform holds, which for an ambient product means the encounter audio question is unanswered. Ask what the agreement permits you to switch off, since the company's own wording implies it can be varied.

The company states it maintains HIPAA compliance and enters into business associate agreements with covered entities, and it does so in its terms of service rather than in marketing copy. Committing to the agreement in the contract document is a firmer position than asserting compliance on a web page.

Held at B rather than A because the agreement itself is not published, so a buyer can see that one exists without seeing what it says about breach notification timing, subprocessors, permitted uses or return of data on termination.

One point to raise given the product's shape. The platform both listens to the encounter and submits against it, so the same vendor holds the audio, the note derived from it and the claim built on it. Establish whether the agreement treats those as one data set with one retention rule or three, because a recording, a note and a claim have very different lifespans and very different exposure if retained.

SOC 2 Type II with the type named and regular third party audits stated, alongside encryption in transit and at rest and customisable role based access. Crucially this is the company's own attestation rather than its cloud provider's, which is the distinction this index most often has to correct.

Where it is published is worth noting. The commitment sits in the terms of service rather than on a marketing page, which makes it a contractual statement rather than a claim, and that is stronger.

Held at B rather than A because this is a single framework with no HITRUST or ISO 27001 alongside it, no trust centre or self serve report access exists, and no observation period or scope is stated. For a platform ingesting encounter audio and clinical notes across multiple electronic health records, ask which systems are in scope and whether the ambient scribe is inside the audited boundary or beside it.

No FDA pathway applies and none is claimed. Provider side coding and revenue cycle work has no vendor level regulator at all, which is the structural finding this index records across the category: liability for what is submitted sits with the practice under the False Claims Act, not with the software.

Graded B rather than C because the company operates against the applicable rule set explicitly rather than leaving it implicit. Its outputs cite the underlying rule and its source from CMS, the national correct coding initiative, or payer policy, and its terms state plainly that it does not provide coding advice and that customer personnel remain responsible for final coding and billing. Naming the rules you answer to and the boundary of your responsibility is the substitute for a clearance in a category that has no regulator.

One tension to press. The service list includes autonomous medical coding while also describing human in the loop review for exceptions, which means non exceptions are coded without a human. Establish what makes something an exception, who sets that threshold, and whether a credentialed coder sees anything that was handled automatically.

No AI governance framework, model monitoring disclosure or bias evaluation was located.

The direction question is the one that matters for a coding product and it is unanswered. A system that reviews every encounter and returns suggested corrections is changing what gets billed, and the useful disclosure is the ratio between codes it adds and codes it removes. Every published figure here points one way, fewer denials and higher accuracy, and nothing reports how often the system recommends removing a code that documentation does not support. Among the vendors this index has examined in this category, only one publishes both directions, and this is the standing question for all of them.

A second mechanism applies to the scribe. The company markets multilingual support and understanding of accents and jargon, which is a claim about performance across speakers. Nothing published tests it. Where a scribe feeds a coding engine, a documentation gap caused by a misheard encounter becomes a billing consequence, and the clinicians and patients affected will not be evenly distributed.

The explainability here is the strongest form found in coding products and it carries the grade. Outputs carry the underlying rule and cite its source, so a coder sees why a change was proposed rather than receiving an unexplained flag, and can evaluate the proposal against the rule it claims to apply. That converts review from acceptance into adjudication, and in autonomous coding it is the control that lets a human refuse the machine on grounds they can articulate to an auditor later.

Against that, the headline number is unsupported in a way that matters more here than elsewhere. A claim of 98 percent coding accuracy carries no audit standard, no denominator, no specialty breakdown and no independent validation, and coding accuracy is meaningless without stating what it was measured against and by whom, since a vendor grading its own output against its own reference can reach almost any figure.

For a product performing autonomous coding the two numbers worth publishing are accuracy against a credentialed coder audit and the rate at which autonomously coded claims are later denied or reversed, and the second is the one a practice actually feels. Neither exists. No warranty, indemnity or remediation commitment was located. Ask for both numbers and for the audit standard behind the published one.

Sits on top of existing systems, pulling clinical notes, codes, and charge files from Epic, Oracle Health, athenahealth, and ModMed. Read breadth across four major systems is solid for a seed stage company; write back depth into discrete fields was not documented.

United States based cloud infrastructure is stated, which is a genuine residency commitment and more than many records here carry. It is credited, and it is the only element of the deployment picture that is published.

No hosting provider, region, tenancy model or subprocessor list was located. For a product combining ambient capture with coding, two subprocessor layers are near certain and neither is disclosed: speech processing for the scribe, and a language model provider for note generation and code suggestion.

That second one is the question to press. Establish where inference executes, whether encounter audio or note content reaches an external model provider, and what that provider retains. The company's own terms state that de identified data improves model accuracy, so a buyer should also establish whether that improvement happens inside the United States boundary the company advertises.

No rate card, but the evaluation path is unusually concrete and low friction: providers can submit a few hundred cases for initial analysis with results in about three days, and a free ICD-10 code search tool is offered publicly. A prospect can test the product against their own data before committing, which is a more useful disclosure than a price for a product whose value depends entirely on the buyer's case mix.

Scope is stated concretely rather than in the abstract. Buyers are named as specialty physician practices, surgery centres and health systems, with orthopaedics, dermatology, ophthalmology and gastroenterology called out specifically, and the scribing component is described as covering more than forty specialties with support for multilingual encounters.

Specialty focus is the right posture for this product, because coding rules, modifier conventions and payer policy differ sharply between specialties and a general coding engine tends to be weakest exactly where the money is.

Held at B rather than A on the gap between the two claims. Four specialties are named for the coding and denial work while more than forty are claimed for scribing, and those are different depths of capability sold under one name. A buyer outside the four should establish what is production proven in their specialty rather than assuming the wider number applies to the revenue integrity side.