Switchboard, MD

The models are the engine of the core module, and a workflow platform sits around them. Proprietary models classify every inbound item (referrals, portal messages, calls, voicemails, faxes, texts, forms), extract details, prioritize, and route each to the right work queue, and conversational voice and text agents handle outreach and scheduling.

Around them the company sells work queues by task type, multi-step outreach campaigns triggered by workflow status, referral and capacity analytics, and routing that can be 'automated, rules-based or AI-driven'. With the models removed, a buyer would still hold the queues, campaign builder and dashboards, but the classification and routing that make them work would be manual.

Work reaches a person, and the company promises urgent items are never missed. The structure: every inbound item is classified and placed in a work queue for the right team member, with routing that can be rules-based or AI-driven, suggested next actions, and voice and text agents that contact patients and book appointments against real-time availability. The promise: the platform page describes 'Context-aware prioritization: Urgent needs surfaced instantly, every time.'

That is a claim about what the model does not miss, which review of the queue cannot show, because a message classified as routine goes to a non-clinical queue and waits. The company's own published study reports 97.8 percent classification accuracy, so some messages, urgent ones included, are misclassified. The models are also described as 'continuous learning based on outcomes', so their behavior changes as teams work.

How the models are built is published; which version runs is not. The company describes proprietary models built by physicians and data scientists, 'trained on millions of real-world clinical data points' and 'not borrowed from generic, off-the-shelf LLMs'.

Its peer reviewed study (NEJM AI, February 2025, Anderson and colleagues from Emory University, Georgia Tech and Switchboard, MD) describes the founding model: a pretrained BERT model fine tuned on clinician labeled patient messages into five classes (urgent, clinician, refill, schedule, form), validated on 500 held back messages.

The company says that model was an early version of its product, which has since evolved, and the platform page describes 'continuous learning based on outcomes: models improve as your team works'. The company does not say which model version is deployed, how it differs from the studied one, or when it changes.

The model layer is stated to be built in-house, which keeps that part of the chain short. The platform page says the company's AI is 'built from scratch by physicians and data scientists' and 'not borrowed from generic, off-the-shelf LLMs', and its published study describes a fine-tuned BERT model; the platform can also be deployed on premises, where patient communication need not leave the customer's environment.

What remains unnamed is everything else in the chain for the cloud deployment and the voice agents: hosting provider, speech recognition and synthesis, telephony and text messaging providers. No subprocessor list is published.

Peer-reviewed, prospective evidence for the message triage model, and vendor figures for everything the site now leads with. The NEJM AI study (February 2025; authors from Emory University, Georgia Tech and Switchboard, MD, including the founder) validated the model at 97.8 percent classification accuracy on 500 held-back messages labeled by clinicians, then deployed it across four Emory outpatient locations, comparing 469 routed patient conversations with 402 unrouted ones from the same period: staff had two fewer message interactions per inquiry and conversations closed 22.5 hours sooner, an 84 percent reduction in median resolution time.

The company says the studied model was an early version of its product. The home page now leads with referral conversion, reporting 30 percent more referrals converting to scheduled appointments, 33 percent higher callback answer rates and 60 percent more capacity, with no customer, period or method; the study does not answer those claims. Kirby Medical Center and Georgia Vision Institute are named customers.

Categorical positions are stated, including the training position, without a retention schedule. The company's announcements say the platform can be deployed on premises or in the cloud and that the customer's data remains the customer's, so an organization can keep patient communication inside its own environment.

The platform page states the training position in the other direction: models are 'trained on millions of real-world clinical data points' and improve continuously 'as your team works', based on outcomes. The website privacy policy covers the site and patient text messaging, telling patients that appointment reminders and alerts by text may carry health information over unencrypted SMS with their consent. No retention period, de-identification method, or statement of whether one customer's data trains models used by others is published.

HIPAA compliance is claimed and business associate status is not stated. The company's announcements describe the platform as HIPAA compliant alongside a SOC 2 Type 2 certification, and the platform page says it is 'secure by design with role-based access and full compliance'.

The only published privacy document is the website privacy policy, which covers the site and patient text messaging, including consent for appointment reminders that may carry health information over unencrypted SMS, and excludes other personal data the company collects. No business associate agreement or statement of business associate status appears on the site.

SOC 2 Type 2 is claimed as the company's own in its announcements, with nothing to open. The company's formal announcements describe the platform as certified to SOC 2 Type 2, and the platform page describes role-based access. The current website carries no security page, trust center, report request route, audit period, auditor, penetration testing statement or subprocessor list.

A Type 2 report tests whether controls operated over a period, so its period and scope are the buyer's first questions, along with whether it covers both the cloud and on-premises deployments and the voice agents added since the study.

The platform sorts patients' messages and calls by clinical urgency, and no regulatory position is stated anywhere the company publishes. The platform page calls it 'a clinical intelligence platform' that 'routes, prioritizes, and actions what matters most', with 'context-aware prioritization: urgent needs surfaced instantly, every time', and the company's study trained the model to place each patient message in one of five classes, including 'urgent' and 'clinician'.

Messages the model classes as refill, scheduling or forms go to non-clinical queues, so the model decides which messages a clinician never reads. No named triage protocol sits behind the urgency class, no clinician decides the messages classed as routine, and the product is not described as a device. Referral capture, scheduling and outreach are administrative.

The model has been tested in public; how it is governed in use is not described. The peer-reviewed study (NEJM AI, 2025) sets out how the triage model was trained, labeled by clinicians and validated on held-back messages before deployment, and the company cites it as evidence its models are 'rigorously tested through academic studies'.

No governance framework, subgroup or bias analysis, drift monitoring, or approval process for model changes is published, and the platform describes continuous learning from team outcomes, so the deployed model changes without a stated control. Triage from free text carries a specific exposure: how a patient writes varies with literacy and first language, and a misrouted urgent message leaves no complaint. The voice agents add accent and speech variation.

A published error rate with its method and denominator, for an earlier version of the model. The company cites its NEJM AI study (February 2025; Anderson and colleagues, Emory University, Georgia Tech and Switchboard, MD) on its platform and newsroom pages: a fine-tuned BERT model classifying patient messages into five classes reached 97.8 percent accuracy on 500 held-back messages labeled by study clinicians (four physicians and a nurse), an error rate of about 2 percent, before prospective deployment at four Emory outpatient locations.

That figure can be tested and contested. The company says the studied model was an early version of its product, so the rate describes that version, not necessarily the one deployed, and nothing is published for the voice agents, referral extraction or outreach. No warranty, remedy or service credit attached to a misclassified or lost item is described.

One record system is named and verifiable, and the integration writes as well as reads. Switchboard, MD joined athenahealth's Partner Community as a Solution Partner in July 2026, with its solution listed in the athenahealth Marketplace; that integration classifies inbound referrals, creates or updates the patient record, and writes confirmed appointments back into athenahealth. The founding model ran inside Emory's record system messaging pools.

The platform page describes integration with record, revenue cycle, ERP, CRM and contact center systems through FHIR, HL7 and real-time data exchange, retrieving records, portal messages and appointment data. No other record system is named, and no production deployment on a named system other than athenahealth customers and Emory is described.

Deployment options are stated; the cloud path is not described. The company's announcements say the platform can be deployed on premises or in the cloud and that the customer's data remains the customer's, which lets an organization keep patient communication inside its own environment. The website privacy policy says personal data submitted through the site and patient messaging is transferred to the company's offices and servers in the United States. No hosting provider, region or tenancy model is named for the cloud deployment, and nothing says whether on-premises installations carry the same models and receive the continuous-learning updates.

No price is published, and the shape of the offer is discoverable. The platform is sold through a demo as referral capture and conversion, work queues for appointment requests, medical records, refills and prior authorizations, voice and text outreach campaigns, and referral and capacity analytics, configured for specialty practices and larger groups 'from small groups to enterprise systems' across the inbox, voicemail, contact center and revenue cycle, and it is available through the athenahealth Marketplace.

On the administrative side, the platform turns inbound work into completed tasks: capturing referrals from fax, voicemail, email, portal and web forms, creating the patient record and driving scheduling; classifying appointment requests, medical records requests, refills and prior authorization items into work queues; and running automated voice and text outreach for scheduling, reminders and recalls.

On the clinical side it triages messages: the platform page describes patient portal messages, calls, voicemails and forms classified and prioritized by urgency and routed to the practice's clinicians and teams. Coverage spans specialty practices (the site's current focus), an academic health system, a critical access hospital and an ophthalmology group, across portal messages, calls, voicemail, fax, text, email, web forms and chat, in the United States. Validation sits behind part of it: the Emory message triage study, and Kirby Medical Center and Georgia Vision Institute as named customers; the referral conversion figures carry no named customer.