Pelica

AI agents perform the main operational work rather than only summarizing data for staff. They prioritize member opportunities, conduct outreach, check payer portals, and create call notes. Role-specific copilots support risk adjustment, quality, pharmacy, provider networks, and care management, all using the same member record. The AI Data Analyst adds natural-language investigation of population data.

Ingestion and shared records are essential supporting infrastructure, but the product's central proposition is that agents carry out routine tasks across those records. Human staff set policy and handle exceptions, while the degree of automated execution can vary by workflow.

The product distinguishes three operating levels: drafts for a person to edit, defined workflows with key-step review, and routine outreach that agents carry through until resolution or escalation. Teams choose policy, and cases outside the rules move to a person with their history attached. Those are substantial oversight controls.

However, a featured HealthCare Partners testimonial describes AI-generated call notes as perfect, an absolute accuracy characterization without a measured note-error rate. Human escalation does not establish error-free documentation. The product therefore combines meaningful operational boundaries with an overbroad output claim, while precise confidence thresholds and recovery procedures remain incompletely specified.

Google Cloud, Anthropic, and OpenAI are identified as language model suppliers for tasks such as summarizing clinical documents and analyzing claims. The platform separates a shared data record, role-specific AI functions, and tools that execute outreach or portal work. These details identify both a model approach and actual suppliers. Specific model families, deployed versions, training lineage, and update schedules are not mapped to individual workflows.

A supplier name does not determine which model generated a note or selected an action. Release evaluation results and routing criteria also remain unclear, limiting comparison of behavior across versions and tasks.

The trust center names Google Cloud, Anthropic, and OpenAI for language model processing and separately lists Vercel, Railway, and Supabase for application infrastructure. Other named suppliers support email, collaboration, code management, and internal operations. This is a substantial view of the service chain, with zero retention stated for model inference.

However, an exhaustive mapping of each workflow to its model and provider is unavailable, and explicit provider training permissions remain unclear. The fact that a company appears on an operational subprocessor list does not mean it receives every category of patient data. Processing scope still differs by supplier and function.

HealthCare Partners reports reducing weekly administrative preparation from about eight hours to 15 minutes and saving more than 2,500 team hours over a year across preparation, outreach, and documentation. The June 2026 methodology explains related production measures, including before-and-after task comparisons and user-reported time savings.

It also defines adherence as the share of eligible members reaching an 80% proportion-of-days-covered threshold, rather than a universal clinical outcome. Several broader figures concern an unnamed flagship IPA or multiple deployments and cannot all be assigned to HealthCare Partners. Independent validation, complete denominators, and isolation of patient outcome effects remain limited.

The standard retention period is the shorter of three years or the customer relationship, with irreversible deletion within 15 days after the customer leaves; nonstandard retention can be arranged for mandatory needs. Model subprocessors are subject to a stated zero-retention commitment after inference. These are concrete lifecycle limits, although explicit training permissions, deidentification practices, and exceptions across the full model chain remain unclear.

Evidence links, action logs, and escalation support review of agent work. Detailed hallucination evaluation and model safety incident procedures are not specified. notification obligations concern personal-data security and do not by themselves define correction of an unsafe AI output.

A business associate agreement is signed at deployment, and the service claims HIPAA compliance for PHI handling. The trust center supplies substantive customer-data commitments covering access, retention, deletion, security incident notification, and a designated privacy officer. Approved personnel can access customer information for support and maintenance under least-privilege controls. These details establish more than a standalone compliance badge.

The actual BAA template and its subcontractor obligations require a request, leaving the complete contractual scope unavailable for direct inspection. The separate website privacy policy does not replace those service terms, particularly for patient records used in agents and outreach.

SOC 2 Type I and Type II reports name MJD Advisors as auditor, with access available by email request. HITRUST is described as in progress rather than an achieved certification. The trust center also describes quarterly access recertification, daily database backups, AES256 storage encryption, TLS1.2 transmission protection, and a 24-hour recovery objective. These controls give useful operational detail alongside the audit claims.

The current reports, examination dates, scope, and exceptions are not downloadable without requesting access. Consequently, the visible assurance does not allow direct confirmation of every current service and model-processing component covered by the audits.

The current offering emphasizes quality reporting, risk adjustment, pharmacy adherence operations, provider preparation, outreach, and care coordination. Agents organize work and documentation around existing member records and customer policies. This is an operational service for organizations responsible for care programs and financial outcomes, without a claimed medical device indication or identified FDA clearance.

References to earlier intervention or better outcomes describe intended benefits rather than regulatory authorization. The choice of automation level also does not establish authority to diagnose or treat. Clinical exceptions are routed to staff within the customer's workflow, preserving a distinction between operational task execution and professional care decisions.

Action logs and evidence links let staff inspect what an agent did and the information behind a recommendation. The outcome methodology defines operational measures and states their limitations. Neither function establishes a formal governance framework for model release, behavioral testing, or demographic fairness. Comparative language recognition results are also unspecified despite multilingual outreach. This leaves limited visibility into how model changes are evaluated for uneven performance across patient groups or languages.

Customer-defined policies, evidence links, action logs, and escalation give staff practical routes to review agent work and correct an operational record. The methodology explicitly states that reported customer outcomes are not guaranteed for a new deployment. A product accuracy warranty, indemnity for an erroneous output, or contractual correction and compensation obligation remains unspecified in the accessible service material.

The website terms are not a substitute for a signed organizational agreement. Thus, the concrete recourse visible to users is review and intervention within the workflow, rather than an assured financial remedy for an incorrect note, outreach decision, or risk adjustment recommendation.

The data layer accepts claims, EHR information through FHIR and HL7, pharmacy data, laboratory results, admission events, and payer file transfers. Epic SMART-on-FHIR and Athena overlays are named, while the HealthCare Partners case supports operational ingestion of payer, medical, laboratory, and pharmacy information. These facts establish more than a generic interoperability promise.

However, the exact bidirectional production scope, supported versions, and writeback behavior for each named EHR remain unclear. An agent updating Pelica's shared member record is not necessarily the same as updating the source clinical record. The strongest evidence concerns intake and workflow access rather than comprehensive EHR exchange.

Customer data is described as hosted exclusively in the United States, with a separate data plane for each customer and encryption in transit and at rest. The infrastructure list names Vercel for frontend hosting, Railway for backend services, and Supabase for database, authentication, and storage. Role-based access and customer-managed keys are also described. These are substantive hosting and isolation details.

The remaining uncertainty is whether the US commitment covers every model inference operation, support path, and downstream processor. Supplier identities and retention terms describe different aspects of processing from the location where inference occurs.

Customers buy an organizational platform deployment with implementation support and a signed service agreement. The service can bring together data ingestion, role-specific copilots, and action tools, with each layer described as usable independently. HealthCare Partners reports a two-week interval from signing to production, but that implementation experience does not establish a standard fee or guaranteed timeline.

Subscription prices, contract values, charging units, and minimum commitments are not provided before direct engagement. The methodology expressly excludes public pricing and contract values. As a result, buyers can identify the service components but cannot calculate the initial or recurring cost from a published schedule.

HealthCare Partners is a named IPA deployment spanning quality, pharmacy, nursing, and risk teams, with approximately 1,000 providers and more than 200,000 members in New York. Its case illustrates coordination across several operational roles rather than a single specialist application. The wider offering addresses Medicare Advantage plans, other health plans, ACOs, management organizations, and medical groups.

Pharmacy adherence, risk capture, transitions of care, and provider preparation give the platform multiple use cases. Equivalent measured results are not established for every advertised organization type or clinical population, and the named IPA's baseline workflows affect how transferable its savings are.