Freed

Ambient capture to structured SOAP note is the entire product. Nothing remains without the model.

Clinician control is stated as a design principle rather than a disclaimer, and it is backed by specific mechanics. The clinician maintains complete control over what is documented and how notes are finalised, the final note can be edited at any time, and notes can be deleted at any time. A quality check step runs before the audio recording is deleted, which implies a verification stage between generation and finalisation. One mechanism is distinctive and cuts both ways.

The system learns from clinician edits, inviting the clinician to have Freed learn a correction for next time, so the oversight step is not only a safety check but a training signal. That is genuinely good for fit, and it raises a question nothing public answers: whether corrections stay local to that clinician's model or propagate to other users, and what happens when a clinician teaches the system something incorrect.

A personalisation loop that learns from unverified human edits has a failure mode a static model does not. Held below a higher grade on the usual absences for this category. There is no published accuracy, hallucination or omission rate, nothing stating what the system does when audio is poor, speakers overlap or a section cannot be heard, and no indication to the reviewing clinician of which parts of a note are low confidence. A reviewer not told where the uncertainty sits is exposed to the automation bias that fluent, usually correct drafts produce.

The training data position is disclosed plainly, and it is the disclosure this index asks every ambient vendor for and rarely gets: Freed states that its AI is trained only on de identified notes, and that protected health information is never used for AI training.

Set that against the rest of the category, where one vendor claims continuous AI training while leaving the source unstated, another claims an extensive dataset from record integrations while separately saying voice is used only for transcription and never reconciling the two, and a third says nothing at all. A clear statement of what the model is and is not trained on is more useful to a buyer than the name of the foundation model.

Two smaller disclosures support it: automatic speech recognition is described as specifically trained for clinical use and deriving medications, dosages and abbreviations, and the AI assistant is said to draw on more than 50 trusted sources. Held below a higher grade because the model itself is not characterised.

There is no model class or version, no foundation model provider named, no architecture description, no update policy, no statement of what is proprietary versus wrapped, and no published accuracy, error or benchmark figure of any kind. Microsoft Azure is named as infrastructure, which identifies the cloud but not the model, and the 50 plus sources behind the assistant are counted but not enumerated.

One of the better disclosures in this lane, and it is built from several partial answers that add up rather than from a single artifact. The cloud is named, with all data processed and stored in the United States on Microsoft Azure, and the business associate agreement with that provider is named as well, which is the detail most vendors omit even when they name the cloud: it tells a buyer not just where content goes but on what legal footing.

Sub processor governance is published as a rule rather than a list, with every vendor that may process patient information required to be compliant and to sign an agreement with Freed. And the training corpus is characterised directly, with a statement that protected health information is never used for training and that the model is trained only on de identified notes. Held below the top grade because none of it is enumerated.

The rule about sub processors is published without the resulting list, so a buyer knows the standard applied without knowing who met it. No foundation model provider is named, no model class or version is given, and the more than fifty sources described as feeding the assistant are counted rather than listed. One point belongs on this axis as well as on stewardship, because it changes what the chain contains.

Training only on de identified notes means customer encounters do form the derived corpus, with de identification as the mitigation rather than an exclusion. That is a legitimate design and it is not the same claim as not training on customer data at all. Ask for the sub processor list and for the sources behind the assistant.

No peer reviewed study, controlled evaluation, published accuracy measurement or quantified time savings result was located from a vendor source. What exists is adoption scale and clinician testimony: use claimed across hundreds of health systems and thousands of clinicians, alongside individual practitioner quotes. Scale of use does not substitute for evidence of benefit, and this index applies that consistently across categories.

One thing is worth crediting because it is a structural alternative to published evidence rather than a substitute for it. A free trial exists, and the company explicitly encourages prospective buyers to run their own pilot and generate their own data.

For a product sold to small community practices that will never appear in a multicentre study, letting the buyer measure it themselves on their own notes is a reasonable answer, and more honest than a vendor generated figure with no denominator. It is not the same as evidence, though, because it puts the evaluation burden on a buyer with no control group and no baseline, and most small practices lack the time or method to run a clean comparison. No dedicated research, evidence or outcomes page was located, so this assessment could change on a later review.

Freed states that protected health information is never used for AI training and that its AI is trained only on de identified notes. Read quickly that sounds categorical, but it is not, and the distinction matters. De identified notes are still derived from customer encounters, so customer content does train the model; de identification is a mitigation rather than an exclusion.

That differs from competitors stating outright that they do not train on user data at all, or that data is never used for secondary purposes such as model training. Those are categorical claims; this one is conditional, which means the control is an opt out rather than an opt in default and the burden sits with the buyer to exercise it.

When comparing vendors on this point, it is worth separating the claim that a company never trains on protected health information from the claim that it never trains on your data. Everything else on this axis is strong and the assessment should not be misread as weakness. Audio is not retained by default, with recordings kept only until the note is generated and quality checks complete, then automatically deleted.

All data is processed and stored within the United States, in Azure, under a business associate agreement with Microsoft. Subprocessor governance is published: every vendor that may process patient information must be HIPAA compliant and sign an agreement with Freed. Encryption is stated precisely, cryptographic modules follow FIPS PUB 140-2, and clinicians can edit and delete notes at any time.

A Business Associate Agreement is provided by default rather than gated behind an upper pricing tier. That is a direct contrast with Heidi and Nabla, where the BAA is reported as unavailable on entry tiers, and it means the published $39 plan is a viable production configuration for a US clinician handling PHI rather than an evaluation toy.

This is the most specific security disclosure in the ambient scribe category. Certifications are named with their types: SOC 2 Type 1 and Type 2, described as independent third party audits, plus HIPAA and HITECH compliance. Three items go well beyond what this category normally publishes. Cryptographic modules follow FIPS PUB 140-2, a United States federal standard almost nothing in this index cites and a far more precise claim than enterprise grade encryption.

OWASP secure coding standards are enforced with regular audits, naming the framework rather than asserting secure development in the abstract. And the monitoring toolchain is named outright, using Azure Security Center and Drata for continuous monitoring and vulnerability scanning. Encryption is stated precisely as TLS 1.2 to 1.3 in transit and AES-256 at rest, with hosting on Microsoft Azure.

Operational controls are documented rather than implied: infrastructure as code with all changes reviewed before deployment, secure development training for every engineer, background checks on all staff, annual HIPAA and privacy training, single sign on and multi factor authentication. Subprocessor governance is the part most vendors skip entirely.

Freed holds a HIPAA compliant business associate agreement with Microsoft, requires every vendor that may process patient information to be HIPAA compliant and sign an agreement with Freed, and states that it reviews vendor security practices on an ongoing basis.

No FDA clearance, none claimed and none required for ambient documentation. The base position matches the category: the system drafts, the clinician reviews, edits and finalises, and the clinician retains complete control over the note. The scope expansion pattern appears here too, and on this record it spans the widest functional range in the category even though the company serves the smallest customers.

The published product set covers ambient scribing, pre charting and visit preparation, ICD-10 and CPT coding suggestions, patient instructions and post visit letters, specialty specific templates, an AI assistant that answers medical questions from more than 50 trusted sources tailored to patient context, and a front desk AI receptionist that answers patient calls. Three separate boundaries are crossed and a buyer should evaluate them separately.

The AI assistant is clinical decision support rather than documentation, landing in the territory occupied by dedicated clinical reference products. The coding feature is marketed as suggesting codes at the highest justified evaluation and management level, and the qualifier is doing real work in that phrase: a system optimising toward the top of a defensible range is a different object from one that codes what happened.

And the front desk product puts AI in direct contact with patients rather than clinicians, which is a different safety posture entirely. Worth establishing which modules are in scope, and asking for the rationale on the assistant and the receptionist specifically.

No AI governance artefact was located: no responsible AI statement, no bias or fairness position, no demographic or subgroup performance analysis, no error taxonomy, no model monitoring description and no published evaluation output. The juxtaposition is the point here. This vendor's security documentation is the most detailed and specific in its category, naming its cryptographic standard, its secure coding framework and its vulnerability scanning toolchain.

A company willing to publish that level of engineering detail, and then publish nothing about how its model performs across different speakers, has made a choice about which kind of trust it is building. Security is treated as an engineering discipline with named standards; model behaviour is not treated at all.

One adjacent signal deserves mention without changing the assessment, because it shows the issue is on the company's radar: the product is marketed as having specialty awareness including an understanding of sensitive conversations across specialties such as behavioural health. Recognising that some encounters are more sensitive than others is a start, but it is a content handling claim rather than a performance disclosure.

The unaddressed exposure is the same one that applies to every ambient product: speech recognition accuracy varies measurably across accent, dialect, speech rate and vocal characteristics, and the affected populations overlap with those already underserved. A vendor serving community clinics has a stronger reason than most to measure and publish it.

The correction machinery is real and it is the reason this sits mid band. Clinicians can edit and delete notes at any time, recordings are kept only until the note is generated and quality checks complete and are then deleted automatically, and the data commitments around them are specific rather than gestural.

Taken together that gives the person who signs the note both the ability to fix what the system got wrong and a bounded window in which the raw material still exists to check it against. What is absent is any commitment about the output itself. No accuracy, error or benchmark figure of any kind was located, no performance guarantee, no remediation obligation and no indemnity toward the customer.

The speech recognition is described as clinically trained and deriving medications, dosages and abbreviations, which is a claim about capability with no published measurement behind it, and medication and dosage extraction is precisely the output where an error carries the most weight. The patient has no route, as throughout this category. Ask for accuracy on medication and dosage extraction specifically, and for what the vendor commits to when the note is wrong.

EHR Push works through a Chrome extension that pastes the note into any browser based EHR with one click. There is no native EHR API and no structured field level write back, so discrete data does not populate the chart. That is a real constraint for any organization wanting coded data in the record, and it is the principal tradeoff of the self serve model. The vendor and third party reviews both state it plainly.

A data residency statement exists and is unambiguous: all data is processed and stored within the United States, in Microsoft Azure data centres, under a HIPAA compliant business associate agreement with Microsoft. Most competitors in this category publish nothing equivalent, and for a product recording patient encounters this is the question a compliance officer asks first.

The hosting provider is named rather than described, and Azure high availability infrastructure is cited for continuity. Delivery is web browser plus mobile application with no additional hardware, and the product explicitly handles offline and low connectivity environments, which is a real deployment property in rural and community settings and is rarely addressed in this category.

Evaluation is genuinely self serve: a free trial exists, so a clinician can deploy and test without an institutional purchase. The target segment is stated with unusual directness, and it is a deliberate exclusion rather than a gap. The company says it is not made for massive healthcare systems and is built for community care, one local clinic at a time. Naming who you are not for is a form of candour worth crediting.

Held below a higher grade on three points: residency is stated at country level with no region, failover location or subprocessor list beyond Microsoft; no customer is named, which the company says is deliberate but leaves adoption unverifiable; and no uptime commitment or support model was located.

The most complete pricing disclosure in the index. Fully published self serve tiers with amounts: Starter $39 per month for up to 40 notes, Core $79 per month unlimited, Premier $119 per month or $104 billed annually adding EHR Push and coding. A seven day trial requires no credit card, students receive 50 percent off, and groups of ten or more get custom quotes. A buyer can determine total cost, compare tiers, and start using the product without ever speaking to sales, which is exactly what this axis rewards.

Fit is strongest in primary care and general practice. Specialty depth outside that is a documented weakness, with procedure heavy and subspecialty documentation reported to need more clinician editing. Coverage is honestly characterized rather than overstated, but the validated range is narrow relative to specialty focused competitors.