Ambient Scribes
B

BastionGPT

BastionGPT is a healthcare AI assistant and medical scribe from FortaTech Security, which trades as Bastion Intelligence. It gives clinicians a private alternative to ChatGPT for protected health information, routing each task among licensed GPT, Claude and Gemini models and covering every plan, trial included, with a business associate agreement.

The scribe records visits of up to 240 minutes with up to 10 identified speakers, joins Zoom, Teams and Google Meet calls without a visible bot, or takes an uploaded recording, and drafts SOAP, DAP, BIRP, history and physical, intake, treatment plan and referral notes. Clinicians can feed it example notes to shape style and structure. The assistant also summarizes records of up to 1,000 pages on higher plans, drafts letters and supports clinical reasoning, with handling for sensitive topics such as trauma, sexuality and substance use that general assistants often refuse. Notes reach Epic, athenahealth, SimplePractice and other systems by copy and paste.

BastionGPT says it serves more than 10,000 healthcare organizations and, by its own 2024 survey, saves clinicians about 90 minutes a day. It cites a Baylor College of Medicine and Texas Children's Hospital study in which 8 of 1,225 generated sentences had minor errors. Plans list Professional at $20, Professional Plus at $45 and Ultra at $75 per user a month, though its scribe page still advertises a $45 starting price.

AI Health Index verifiedNovember 19, 2026
Compare BastionGPT with other vendors
Founded
—
Headquarters
—
Website
bastiongpt.com
Categories
ambient-scribes, clinical-summarization
Indexed Products
BastionGPT, AI Medical Scribe, Professional Plus, Enterprise Ultra
Buyer Segments
Clinician, Practice, Health System, Behavioral Health
Assessment

Capability Axes

An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read

AI Capability
AA on AI CentralityThe artificial intelligence is the product. Remove the model and there is nothing left to sell.
Vendor Published

Everything BastionGPT sells runs on language and speech models: chat, medical transcription, note generation and synthesis of clinical records. Its healthcare setup handles sensitive discussions about trauma, sexuality and substance use that can trigger refusals in general assistants.

Clinicians supply a preferred note or template and get SOAP, DAP, BIRP, referral and other documents from the encounter, and model routing and clinical evaluation are part of the service rather than extras bolted onto a record system. Document storage, account administration and template selection are ordinary software, but the product's main deliverables depend on learned models.

BB on Autonomy and Oversight ModelThe oversight structure is described and one part is missing, commonly the threshold at which the system stops or what happens after it is wrong.
Vendor Published

Clinicians read the transcript and the generated draft before copying anything into their own record system. Saved instructions and sample notes shape the output, and higher plans let users pick the model by hand. The vendor's evaluation treats invented details, omitted risk and overreach into clinical authority as failures, and requires human judgment for clinical use.

There is no signature gate inside the patient's EHR, and the assistant can still give wrong or incomplete answers. In the pediatric ICU study, parents talked with it unsupervised for 10 minutes before an investigator went over the transcript with them.

BB on Model and Technology TransparencyThe approach or the suppliers are named without the version and update discipline behind them. Naming a supplier is the entry to this band both here and on Model Supply Chain Disclosure, which ask different questions of the same disclosure: who receives the data, and what produces the output.
Vendor Published

The assistant routes work among licensed OpenAI, Anthropic and Google models. A July 2026 evaluation names GPT-5.5, Claude Sonnet 4.6, Gemini 3.5 Flash and Gemini 3.1 Pro, run on the same prompts across 40 synthetic cases, and performance by task family decides which models handle which work.

An internal fallback model handles legitimate clinical material the outside models refuse, and its architecture, training data and version are not given. Scribing adds speech recognition and acoustic speaker separation before the note is written. The 2024 ICU deployment used GPT-4o with little healthcare tuning, a different setup from today's multi model service.

BB on Model Supply Chain DisclosureSubstantial partial disclosure, or a chain that is structurally short: an in house build, a cleared model that cannot be quietly swapped, or a deployment where the transfer does not occur at all. Naming the model provider exits the band below into this one; the axis rises from here on the completeness of the party list and on the terms that govern data once it arrives.
Vendor Published

Microsoft and Google are the named providers that receive customer content, for cloud storage, speech, document processing and model inference. The supplier list names Anthropic only for internal business use without customer content, even though Claude models are offered in the assistant, so licensing a model does not mean patient data goes to its developer.

Payment, marketing, scheduling and customer communication providers get no customer content. Contracts bar providers from training on submitted content and pass equal data protection duties down the chain. Which hosting route each model and the internal fallback take is not stated.

BB on Clinical and Operational EvidenceNamed deployments with dated outcome figures and enough method to test them, or published research short of independent validation.
Vendor Published

A July 2026 BastionGPT evaluation ran GPT-5.5, Claude Sonnet 4.6, Gemini 3.5 Flash and Gemini 3.1 Pro on 40 synthetic clinical cases each, scored by an AI judge against case specific criteria with clinician spot checks. Of 160 raw model responses, one crossed a critical failure line and nine were flagged for clinician review. The tests used identical prompts without the product's own system prompts, so they do not measure the scribe end to end.

The site markets a peer reviewed pediatric ICU study as 99.3% accuracy on documentation tasks. That study measured eight minor errors in 1,225 sentences GPT-4o wrote answering parents' questions, in English only, which says little about today's scribe.

AA on AI Safety and PHI StewardshipRetention windows, training use and de identification are all stated specifically enough to be contradicted. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, does not gate this grade: no vendor in this index publishes it, so every note on this axis states that gap rather than leaving it implied.
Vendor Published

Prompts, uploads, audio, transcripts and outputs cannot be used to train, fine tune, evaluate or improve future models, and the bar extends to deidentified or aggregated versions of that content. After termination, PHI can be retrieved for 30 days, requested exports arrive within 15 days, and destruction follows within 30 days after that, with backups purged within one further backup cycle.

If destruction is not feasible, BastionGPT must say so in writing and keep protecting the data. Encryption runs on AES-256 at rest and TLS 1.3 in transit. None of this stops a model from reasoning badly, so clinical output still needs review.

Regulatory and Compliance
AA on HIPAA and BAA PostureBusiness associate status is stated, the agreement is available to read before contracting, the tier it applies at is clear, and the agreement discloses that it binds subcontractors who handle protected health information. Which parties those subcontractors are is graded on Model Supply Chain Disclosure and is not counted again here. Where any of the four cannot be verified from published material, the grade is the band below.
Vendor Published

The business associate agreement takes effect when a covered entity or business associate accepts the service terms and first sends PHI, and it applies to every plan and every trial with no separate signature. Its channel list covers signed in chat, uploads, transcription, generated outputs, API access and beta features, and excludes ordinary support email, billing fields, marketing forms and community channels unless the company directs a secure transfer.

Subcontractors handling PHI accept equal restrictions, a Part 2 addendum applies automatically to substance use records, and PHI sent through an excluded channel is removed with written confirmation.

CC on Security Certifications and Trust CenterControls are described with an outside check behind them, such as independent penetration testing on a stated cadence, but no attestation against a recognized framework. A certification that tests the software security features rather than the vendor operations, such as federal health IT certification, is an outside check of this kind.
Vendor Published

Controls include encryption, tenant separation, access restrictions, monitoring, incident response and annual independent penetration testing that reaches application code and AI threats such as prompt injection and data exfiltration. Enterprise plans add single sign on and custom security arrangements, and the site shows Qualys and SecurityScorecard badges.

Bastion's own HITRUST and SOC 2 work is at the readiness stage, so the completed certifications in the picture belong to its cloud providers and cover their infrastructure. Assessment reports are available through a controlled trust process, and no certification issued to BastionGPT itself, with a period and boundary, is in place.

CC on FDA and Regulatory StatusNo device claim is made and the product is scoped accordingly. Most administrative and operational products sit here and are not penalized for it, because this axis grades the appropriateness of the positioning rather than possession of a clearance. A stated limit on use, such as research use only or a clause that the product gives no medical advice, beside marketing that sells the product for clinical decisions also sits here, with both quoted, because the buyer has to ask which one governs.
Vendor Published

The contract defines the assistant as a professional support tool, not a medical device or a source of medical advice. In practice it goes beyond formatting into differential reasoning, treatment plan drafts, medication discussions and patient education, so where it sits depends on the use and the clinician's role.

No FDA authorization is cited for the assistant or the scribe. Its HIPAA, FERPA and substance use confidentiality commitments concern information handling rather than device status, and the institutional research approval behind the ICU study does not authorize the commercial product for clinical decision making.

DD on AI Governance and Bias DisclosureNothing published on how model behavior is governed or tested. Multilingual operation with no subgroup performance sits here when the vendor markets recognition quality as a strength, because a caller the system failed to understand leaves no complaint and no record.
Vendor Published

BastionGPT's clinical evaluation separates the case author, the candidate model and an AI judge, adds clinician spot checks and sets explicit critical failure rules. It tests authority boundaries, unsafe dosing, misleading citations, stigma and prompt injection, and the results decide which models are eligible and how tasks route.

The scribe is also marketed on handling varied accents, languages and speaking styles, but error rates by language or accent are not reported. The ICU study enrolled only English speaking parents and noted underrepresented groups, and neither it nor the synthetic pilot shows that transcription performs evenly across language groups.

CC on AI Liability and RecourseMechanisms exist that let someone challenge an output, such as audit trails, source traceability or review before commit, with nothing standing behind the output and no route for the harmed party.
Vendor Published

Clinicians can read, edit and reject drafts before they reach a patient record, and the model pilot catches failures against defined synthetic case criteria without giving an error rate for the full service.

Contract remedies cover subscription satisfaction and availability: a 45 day money back guarantee and uptime credits, neither of which reaches harm from a wrong clinical output. The terms disclaim accuracy warranties and cap liability at the greater of 12 months of fees or $100, and users stay responsible for how they use what the assistant produces.

Integration and Deployment
DD on EHR and Interoperability DepthNo integration evidence. A connector described as available on request grades here until one exists.
Vendor Published

Notes reach an EHR by copy and paste. The Epic, Cerner, athenahealth, eClinicalWorks, SimplePractice, TherapyNotes and 50 or more other systems shown as compatible describe that workflow rather than installed interfaces, which is why the site can claim no setup steps. The scribe captures live visits, joins Zoom, Teams and Google Meet calls without a visible bot, or takes recordings while clinicians keep their usual telehealth software.

API access exists for selected uses, but no production EHR connection with defined read and write scope is named, and the ICU study entered progress notes, results and medications by hand.

BB on Deployment Model and Data ResidencyOptions and residency are stated with isolation or the processing path left open.
Vendor Published

Standard access is a hosted app, and the account's registration country sets where data lives: US users get US processing and storage, and Canadian and Australian accounts get their own national locations, in line with PIPEDA and Australian privacy principles. Microsoft and Google supply the cloud services, and Enterprise Ultra adds the option of the customer's own enterprise cloud and custom security arrangements.

The privacy terms still allow processing in other countries without mapping support or administrative work to a region. Named cloud regions, failover locations and the edges of each enterprise option are not given.

Commercial
DD on Commercial TransparencyNothing a buyer can establish before a sales conversation. A published pricing claim contradicted by evidence also grades here.
Vendor Published

The plan table lists Professional at $20 per user a month, Professional Plus at $45 and Ultra at $75 for organizations of 100 users or more, with chat, scribing, transcription and a business associate agreement in all three. The scribe page, though, still says the scribe starts at $45 a month while its own FAQ says plans start at $20, so the entry price depends on which line a buyer reads.

Plus expands documents and model controls, and single sign on and negotiated agreement changes are paid extras for Plus groups of 50 or more and included in Ultra. Trials run seven days, and unlimited use is subject to fair use limits.

CC on Setting and Specialty CoverageCoverage is claimed broadly without specifics, or stated clearly with nothing validating it yet.
Vendor Published

Documentation templates cover primary care, behavioral health, psychiatry, pediatrics, internal medicine, urgent care and specialty consults, from SOAP, DAP and BIRP notes to histories and physicals, intake assessments, treatment plans and referral letters. Sessions of up to 240 minutes with up to 10 speakers suit caregiver conversations and group sessions as well as single visits.

The logo row shows clinicians at Mayo Clinic, Cleveland Clinic, Kaiser Permanente and other systems without claiming contracts with them. That breadth comes from configurable workflows rather than validated specialty models, and the one detailed patient study covers a single pediatric ICU.

Alternatives

Alternatives to BastionGPT

Other products that do the same kind of work as BastionGPT, and what each one does.

Wavo Health is an AI scribe and clinical workspace sold to clinicians and independent practices in the United States and Canada by Wavo AI Technologies Inc., a Canadian federal corporation.

AutoNotes is an AI clinical workspace for behavioral health clinicians in the United States and Canada, used by more than 81,000 clinicians who have generated more than 9 million notes.

Comprehend Health, also sold as Comprehend PT, is an AI scribe built for physical therapy. It listens to a patient visit, or to a short summary the therapist records afterward, and writes the daily note, evaluation or progress note in the therapist's own style.

Medical Scribe, sold at medicalscribe.app by Leveled Platforms, Inc., is an AI scribe that records a patient visit or a clinician's dictation and turns it into a note ready for the record.

Medscape Scribe is a free ambient documentation tool launched in October 2024 by Medscape, the clinical information service, available inside the Medscape mobile app and website that a very large share of United States physicians already use daily.

OneChart is an ambient AI scribe from a small Canadian seed stage team in Hamilton, Ontario. The scribe listens during the encounter, generates structured SOAP, DAP, BIRP and specialty notes, and returns them for clinician review and approval before export.

Commercial

Pricing

Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.

Entry Price Pricing Basis BAA Tier Implementation Source
Conflicting starting prices: $20 or $45 per user per month
The plan table sets Professional at $20 per user per month, Professional Plus at $45 and Ultra at $75 with a 100 user minimum. Chat, transcription, scribing and a business associate agreement are included across plans. The scribe page still gives a $45 starting price. Included in all plans and trials. Signed in chat, document uploads, audio, outputs, API access and beta features are covered; ordinary business and support channels are excluded unless the company directs a secure transfer. No standard setup fee is listed. Single sign on and negotiated business associate agreement changes are paid options for Plus organizations with at least 50 users and included in Ultra. Custom security and cloud arrangements depend on the enterprise agreement. Vendor Published

Professional handles a single PDF, Word or text document up to 30 pages. Plus adds multiple documents up to 1,000 pages, larger context, manual model selection, images and more file types. The seven day trial requires payment details and converts to a paid subscription unless canceled. Unlimited use is subject to fair use terms that permit rate limits, alternative pricing, excess use charges or suspension. Refunds can be requested within 45 days of purchase or renewal, and annual prorated refunds carry a 5% fee.