Paratus Health

Generative voice agents are the service's central interface for patients, rather than an optional analytic feature added to conventional practice software. The agents interpret calls, carry conversations through scheduling or intake tasks, and help produce structured records for staff. Outbound workflows similarly depend on language interaction to reconnect with patients and capture responses.

Telephony, scheduling connections, and administrative rules support those functions, but they do not replace the conversational model's role. Removing that capability would substantially change the primary service offered to practices, even though clinicians remain responsible for any medical decisions informed by a call.

The agents can answer, route, and complete routine administrative conversations, with clinic staff involved in setup and ongoing feedback. Current terms acknowledge that AI may need oversight and prohibit sole reliance on its outputs for clinical decisions. Those limits coexist with the homepage's unqualified promise to handle every conversation from beginning to end.

The service does not explain failed interaction thresholds, uncertain intent handling, or mandatory escalation boundaries in enough detail to reconcile that absolute claim. Human responsibility remains explicit, while the product's description of autonomous completion leaves the practical limits of the agent unclear.

The service uses generative AI for natural language calls, intake, and structured clinical documentation. Its privacy policy specifically names OpenAI APIs among processing services, which identifies a model supplier beyond a generic AI description. The terms also distinguish speech and documentation functions from the clinician's decision responsibilities.

Specific deployed models, version identifiers, training populations, evaluation thresholds, and change approval practices remain unspecified. Consequently, a buyer can identify part of the technical approach and supply relationship, but cannot connect the current agent's behavior to a defined model release or reproduce its performance claims.

The service privacy policy identifies OpenAI, Google Cloud, and AWS, plus EHR providers including Athenahealth and Elation. RingCentral is referenced in the messaging provisions. This is a useful partial list of organizations involved in model processing, hosting, clinical data exchange, or communications. The list is expressly nonexhaustive and does not map each workflow to its processors.

Supplier specific retention periods, training permissions, regional processing commitments, and a complete current subprocessor register remain unspecified. Naming a model provider therefore improves visibility without establishing the contractual protections and data path for every downstream service.

The Clove Dental example identifies nine clinics and reports 20% more booked appointments within 60 days, together with estimated annual savings and additional visits. It supplies a named customer and a duration, but no calendar dates, underlying appointment counts, comparison design, or evaluation procedure. The savings and revenue figures also combine observed claims with annual estimates.

Other accounts identify roles or specialties without naming the organizations, and some homepage testimonials are described as inspired by clinics. These statements suggest operational use without establishing a reproducible current evaluation or an independent assessment of call quality and patient outcomes.

PHI is retained for at least six years under the service terms. That wording establishes a minimum retention period rather than a maximum or a promise of deletion at six years. The privacy policy also permits information use for service improvement. Specific model training restrictions, deidentification conditions, deletion of downstream copies, and technical handling of unsafe or fabricated responses remain unspecified. Clinic oversight and activity logs provide a way to review interactions, but they do not complete the account of how sensitive information is reused or how the vendor contains an AI safety incident.

The terms identify Paratus as a business associate when it handles PHI for covered entities and state that BAAs apply to covered entities and subcontractors. The service privacy policy covers clinical information such as visit notes, medications, insurance details, and EHR data, rather than only website browsing. It also describes access restrictions, encryption, and permitted sharing for service delivery. These commitments support a substantive healthcare privacy posture. The actual BAA, its coverage of individual features, and any distinctions between trial and production use remain unspecified in the accessible materials.

Service controls include encryption during transmission and storage, separate tenant data, and restricted staff access according to role. They address practical risks in a system that handles patient conversations and clinical records. Current independent assurance, the scope of any audit, and an accessible certification artifact remain unverified. These operating controls support a security description, but they do not establish a current externally assessed certification covering the service and its model processing.

Current answering-service language explicitly excludes diagnosis and treatment and positions the agent as a digital receptionist for communication and administrative work. Scheduling, intake, insurance checks, and call routing are the main current functions. The terms also reserve medical decisions and patient care to healthcare professionals. This provides an express scope boundary for the present offering, with no device clearance claimed for those administrative tasks.

Earlier descriptions of broader clinical ambitions do not establish the intended use or regulatory status of the current release. The current boundary rests on the stated functions and prohibitions rather than a clinical authorization.

Implementation includes a limited pilot, feedback from clinic staff, adjustment using real calls, and ongoing performance monitoring. This gives customers a practical role in configuring the service and identifying workflow problems before wider rollout. It does not specify a formal AI governance framework, accountable release review body, measurable approval criteria, or independent behavioral audit.

Current patient subgroup comparisons and customer governance review materials remain unspecified. The process therefore describes operational iteration without establishing systematic oversight of model changes or bias.

Activity logs let clinic staff review interactions, and the terms require healthcare professionals to retain clinical responsibility. The liability provisions disclaim responsibility for medical decisions, outcomes, and harm from reliance on AI, with total liability capped at fees paid during the preceding twelve months. Those provisions allocate risk but do not create an output guarantee.

A binding correction obligation, reimbursement for erroneous agent actions, or a defined output dispute process remains unspecified. The practical route described for problems is therefore staff review and ordinary support, without a specific remedy tied to the agent's accuracy.

The terms name Athenahealth, eClinicalWorks, and OpenDental and authorize access, transmission, and processing of PHI for agreed tasks. The privacy policy additionally lists Elation. However, the implementation FAQ says integrations with widely used EHRs are still being built and describes secure exports and API connections as alternatives. These statements leave current connector availability and data direction unresolved.

A named production deployment demonstrating a particular read and write workflow is also unspecified. The legal permission to exchange records therefore supports an integration intention without establishing that each named platform already has an operating connection.

Google Cloud and AWS appear among processing services, and the privacy policy describes tenant isolation and permissions based on staff roles. These are useful architectural controls for keeping organizations' data separate. Its international transfer language says data may be stored or processed in the United States, which is permission rather than an exclusive residency commitment.

Specific hosting regions, model inference locations, backup geography, and cross-border support access remain unspecified. The service therefore identifies parts of its hosted infrastructure and separation model without defining the complete location of patient data throughout processing and storage.

Fees are set in a subscription or service agreement, and the buying path begins with a consultation or demonstration. The terms explain that nonpayment can lead to service suspension but do not provide a current base subscription, charging unit, included call allowance, or implementation rate. Customer savings and payback examples concern the value of particular deployments. They cannot establish a generally available entry price or the total cost for a clinic's expected call volume.

The service targets outpatient communication across dental, pediatric, obstetric, orthopedic, and other specialty practices. Specialty specific call flows can reflect different locations, hours, providers, and appointment needs. Clove Dental is a named example spanning nine clinics, while other customer accounts identify a specialty or organizational role without a verifiable organization name. The Texas answering offering also addresses multisite practices and hospital phone access.

This demonstrates intended breadth, but comparable performance across specialties, languages, age groups, and clinical populations remains unspecified. A configurable workflow is not itself validation that the agent handles each population equally well.