DeepScribe

The scribe is the model. DeepScribe records the visit, transcribes it and writes a structured specialty note, and the same engine powers AI Coding, which places E/M, ICD-10 and HCC suggestions in the note, and SmartPrep, which assembles a pre visit summary for oncology. Customization Studio learns each clinician's preferred structure and phrasing. Nothing in the product works without the models.

DeepScribe drafts and the clinician signs. Notes arrive in the EHR, whether Epic, OncoEMR or another integrated record, for the clinician to review and sign off, and coding suggestions arrive with them. Capture is fully ambient, with no dictation or wake words, so review at signing is the one point where a person checks the output. DeepScribe describes the scribe as fully automated, with no human quality assurance step between the model and the clinician.

The terms put the checking on the customer, who is responsible for confirming the accuracy and reliability of every recording, including those uploaded to the EHR.

Low confidence passages are not marked in a note. A clinician quoted on DeepScribe's homepage says that 90 to 95 percent of the time the note needs no change, which is the routine in which an unmarked error is easiest to sign.

OpenAI's GPT-4 sits inside the documentation engine, by DeepScribe's own account, working alongside its own domain specific models and data. The description comes from a benchmark write up on its site that compared the engine with GPT-4 alone on de identified patient conversations with background noise, long transcripts and complex reasoning, and reported the engine as 59 percent more clinically accurate on critical defects. A February 2024 announcement names AWS SageMaker and Bedrock as the systems behind continuous training and monitoring.

Those disclosures name the approach rather than the current build: the training data is described by size, millions of clinical encounters, rather than by source, and whether GPT-4 is still part of the engine is not said.

The trust center lists four subprocessors. Amazon Web Services hosts the platform and stores application and customer data, MongoDB Atlas holds structured application data, Auth0 handles sign in and single sign on, and Intercom carries customer support messages.

The model layer is named elsewhere. DeepScribe's benchmark write up says its engine uses OpenAI's GPT-4 alongside its own models, and a 2024 announcement names AWS Bedrock, which fronts models from several providers, for training and monitoring. Neither OpenAI nor any other model provider appears on the subprocessor list, so which company's model processes encounter content today, and under what retention and training terms, is not stated.

The business associate agreement requires every agent and subcontractor with access to protected health information to accept the same restrictions as DeepScribe.

Two kinds of outside evidence support the product. KLAS gave DeepScribe a score of 98.8 in its 2025 Spotlight report, built from customer feedback, with 15 A+ marks across culture, loyalty, operations, product, relationship and value. And an August 2025 study in the Journal of Clinical Pathways on DeepScribe integrated in the EHR in oncology reported 16 percent more diagnoses captured per visit, comorbidities recorded 22 percent more often, a 45 percent increase in documented social determinants of health and a 21 percent increase in ICD-10 specificity, with 98 percent of physicians adopting the tool.

Customer figures add scale. Ochsner Health, which rolled DeepScribe out across 46 locations from July 2024, reports 80 percent clinical adoption and a 75 percent cut in documentation time, and Texas Oncology reports an average chart closure time of 1.6 minutes. Both are customer reported figures without a published method.

The commitments on patient data are contractual. The business associate agreement governs protected health information, and DeepScribe describes encryption in transit and at rest, role based access, multifactor authentication and de identification of patient data, with a HIPAA audit report listed in its trust center.

What happens to data afterward is broadly permitted. The terms let DeepScribe create de identified data during the contract and keep, share and use it after it ends for any legal purpose, including improving the service, and DeepScribe owns usage data that cannot identify individuals. The privacy policy keeps personal data for as long as its purpose requires or the law demands, which may extend past the end of the relationship. A 2024 announcement describes continuous training and monitoring, and whether customer encounters, in de identified form, feed it is not said.

The business associate agreement is public in full, as Exhibit A to DeepScribe's terms of use, version 2.1 updated 17 October 2024. In it DeepScribe states that it becomes a business associate on receiving protected health information for or on behalf of the covered entity.

The agreement attaches to every customer. DeepScribe sells only through executed order forms, each subject to the terms, with no free or self serve tier that could fall outside it, and section 2.2 requires any affiliate added to a contract to countersign the agreement without changes. Every agent and subcontractor with access to protected health information must accept the same restrictions as DeepScribe.

The obligations are specific: use limited to the minimum necessary, safeguards under Subpart C of 45 C.F.R. Part 164, breach notice within ten business days with a defined list of contents, other improper uses reported within fifteen business days, amendment and accounting requests answered within fifteen and thirty days, return or destruction of protected health information when the contract ends, and confirmation that the data remains the covered entity's property. The agreement also lets DeepScribe de identify any protected health information it receives, after which the data falls outside it.

DeepScribe's trust center, opened in July 2025, lists SOC 2 Type II and HIPAA. The SOC 2 Type II report covers 1 January to 31 December 2024, and a bridge letter dated 3 October 2025 states that the same controls remained in place and operating through the end of 2025. The report, a HIPAA audit report and the security policies are released on request.

Controls are listed by family: identity and access, security governance with board oversight, third party risk with business associate agreements required of vendors, tested continuity and disaster recovery plans, incident response, cryptography and cyber insurance. The October 2025 bridge letter revised one control to cover AWS boundary protections such as WAF rules and security groups.

No report covering 2025 is listed yet, and no ISO 27001 or HITRUST certification is named.

DeepScribe claims no FDA clearance and states its position in the contract. Under section 2.4 of the terms the customer acknowledges that DeepScribe facilitates the transcription of live interactions and is not a medical device. Section 7.4, a medical services disclaimer, says DeepScribe does not practice medicine and does not make diagnostic, treatment, coding, billing or other clinical decisions or recommendations, and that the professional duty to the patient rests with the customer.

The product sells coding all the same. AI Coding is a headline module with its own pages for E/M, ICD-10 and HCC, captures oncology specific codes in the note and syncs the suggestions into Epic, while the terms disclaim coding recommendations by name. HCC codes drive risk adjusted payment, so the coding module is where the gap between the terms and the product carries the most weight.

A Trust and Safety Suite, launched in February 2024 with its own page, frames DeepScribe's governance around system safety, transparency, data privacy and human centric values. It describes continuous training and monitoring for accuracy and bias mitigation using AWS SageMaker and Bedrock, and an ongoing collaboration with Covenant HealthCare to keep the product aligned with clinical needs. The trust center adds patient FAQ sheets that clinics can hand out.

No evaluation is published to support the bias commitment. There are no results by accent, dialect, language or speaker group, no error taxonomy and no stated monitoring threshold, although the scribe supports conversations in more than 110 languages.

The terms make DeepScribe answerable for the service working as documented and not for the note. Section 7.1 warrants that the service will be provided in a professional and workmanlike manner and conform in all material respects to its documentation, with DeepScribe correcting any nonconformity at its own expense. Section 7.3 says DeepScribe does not warrant that recordings or information will be accurate, relevant, complete or up to date, and makes the customer solely responsible for verifying accuracy and for clinical decisions.

The money terms are tiered. General liability is capped at twelve months of fees, the cap rises to the greater of $1 million or three times twelve months of fees for breaches of confidentiality or of the business associate agreement, and it falls away for fraud, intentional misconduct and indemnity obligations. Section 9.1 has DeepScribe indemnify the customer for its own fraud or misconduct and for US intellectual property claims, while the customer indemnifies DeepScribe against third party claims tied to the care it delivers with the service.

No accuracy commitment, remediation duty or route for the patient is attached to the output.

Integration is two way and runs deepest where oncology lives. In Epic, including Haiku and Canto, a clinician picks the patient from a synced schedule, DeepScribe pulls prior context into the note, and the finished note and coding suggestions sync back in seconds. In Flatiron's OncoEMR, notes populate the patient chart for review and sign off under a January 2025 partnership covering more than 4,200 providers at over 800 community cancer care locations, and an iKnowMed Generation 2 integration serves practices on that system.

DeepScribe's own guides describe two way integrations with athenahealth and NextGen, built on direct API access, that read medications, labs, prior assessments and orders before the visit and write documentation into discrete fields after it. Ochsner Health runs DeepScribe on Epic across 46 locations.

The service is hosted on Amazon Web Services, with structured application data in MongoDB Atlas and sign in through Auth0, according to DeepScribe's trust center. Clinicians use it through a mobile app, inside Epic Haiku, and inside OncoEMR and iKnowMed, and the product is listed on the AWS Marketplace. The privacy policy allows personal data to be kept beyond the end of the relationship.

The region or country where the data rests is not stated.

Buying runs through a demo and an order form. DeepScribe's site offers a demo request and a sign in, its terms place fees, usage limits and other commercial terms in each order form, and it lists no rate, tier, unit of charge or trial.

The return side is quantified on the site. Texas Oncology reports an average chart closure time of 1.6 minutes, Ochsner Health a 75 percent cut in documentation time, and the coding pages describe capturing more diagnoses with greater specificity. What it costs to get there is set in a sales conversation.

Complex specialty visits are the design target, and oncology comes first. The company says it is used across 90 percent of community oncology centers, partners with Flatiron's network of more than 4,200 providers, integrates with OncoEMR and iKnowMed, and sells SmartPrep as an oncology pre visit product. It also publishes pages for cardiology, gastroenterology, neurology, orthopedics and urology, and a partnership with Pearl Health takes it to value based care providers.

Conversations are supported in more than 110 languages. The company positions the product for long, complex specialty visits rather than short, low acuity ones.