Courier Health

AI powers intake classification, patient summaries, recommended actions and administrative agents within a broader patient experience platform. Those functions depend on interpreting records and their context, while the underlying system also supplies case management, communications, consent controls, workflow configuration and reporting. The learned component is therefore central to important modules rather than the entirety of the product.

Patient services teams can use the shared record and coordination tools independently of a generated summary. The combination matters operationally: AI receives context from the platform, and its outputs return to the same workflows where people manage patient access.

AI Intake sorts and classifies forms, extracts information and matches it to existing patient, provider and payer records. Uncertain fields and ambiguous matches are flagged for human resolution, and new patients require confirmation before creation. When information is missing, an agent drafts a provider request for a person to review and send. Across the AI offer, outputs are linked to sources and people decide what action to take.

These steps describe both permitted automation and conditions that require intervention. Claims that forms never sit idle concern processing speed, while the explicit review paths acknowledge uncertainty in the actual output.

The architecture connects patient, provider, payer and program data so AI operates with therapy specific context. It distinguishes embedded assistance, such as summaries, from agents that execute routine intake or appeal steps before human approval. That explains how the models are situated within the product. It does not name the model families, suppliers or versions that perform extraction, summarization and recommendation.

Training data descriptions are limited, and no release schedule or model change notification process is specified. A connected data foundation can improve context without revealing the underlying model design or the validation of a particular model release.

Large language models are part of the AI description, with a stated commitment against using customer data for their training. The current public architecture does not enumerate model providers, inference services or a product subprocessor list. A hosting provider and the technical relationship between the platform environment and external inference services are also unspecified for the current offer.

The data environment claim narrows the intended handling of information without identifying every organization involved. Provider specific retention rules, training terms and the allocation of processing across extraction, summaries and agents remain unknown.

Madrigal is a named customer with reported improvements of 20% in time to therapy and 50% in specialty pharmacy inquiry turnaround. The case explains the move to shared records, secure communication and recommended actions, including a specific example of coordinating a corrected prescription. It does not provide patient counts, measurement dates or a comparison design that isolates the contribution of AI.

Broader platform claims of faster therapy access, more starts and improved persistence are also aggregate customer statements. These accounts substantiate operational use and reported benefit, while leaving the magnitude and causal contribution of individual AI features uncertain.

The AI design states that customer data remains in the environment and is not used to train large language models. Summaries are constrained by user access and patient consent, while uncertain intake fields are sent for review. These commitments address inappropriate reuse, excess disclosure and some extraction errors. The broader statement that data is never stored in connection with AI does not define retention for the patient records, activity logs or outputs that the platform manages.

Numerical deletion schedules, backup handling and the technical boundary around model processing are not specified. The available assurances are substantive but stop short of a complete data lifecycle description.

The current AI platform is described as HIPAA compliant, and its access design incorporates user roles, geography and patient consent. PHI can be obscured when consent or permissions do not support visibility, including within AI summaries. Those are specific product controls. The public privacy policy mainly addresses website interactions, communications and marketing, without establishing a service BAA or availability across contract tiers.

The service agreement's detailed obligations and availability across individual programs remain unknown. HIPAA positioning is therefore clear at the product level while the contractual allocation of PHI responsibilities is not established in the public terms.

The current AI page states that the entire Courier Health platform is SOC 2 Type II compliant. This is a named assurance claim that extends beyond a generic secure hosting statement. The public AI page does not identify the auditor, assessment period or detailed system boundary, leaving the specific terms and report scope unknown. HIPAA statements and consent controls add healthcare context but do not substitute for the SOC report. The available evidence supports a declared independent assurance posture, while leaving the exact assessed services and the coverage of current AI features to the private assessment materials.

The platform supports patient enrollment, therapy access, reimbursement coordination and adherence operations. AI summaries and recommended actions organize the information available to patient services staff, while intake agents prepare records and requests for review. The Madrigal example concerns communicating a prescription correction needed by a specialty pharmacy, rather than the software independently choosing a dose.

No FDA clearance is attached to Courier Health's administrative platform. References to an approved drug describe the customer's therapy, not approval of this software. The current positioning is consistent with operational support whose clinical decisions remain with the treating professionals.

AI features are evaluated against real cases, outputs remain source verifiable and patient consent limits what each user can see. The platform also records engagement and outcome information to support program improvement. These are identifiable governance practices, but the evaluation sets, acceptance criteria and measured results are not supplied. Subgroup testing across language, demographic, disease and access populations is also unspecified.

SOC 2 assurance addresses security controls rather than model fairness. The current account supports a review and evaluation process without showing how model updates are approved, how regression is detected or how unequal performance is corrected.

Every document, status change and interaction is recorded, and AI outputs include sources that a person can verify. Intake requests remain drafts until staff review and send them, creating an opportunity to correct errors before external action. These mechanisms help teams challenge an extraction, interpretation or suggested next step. They do not establish a financial remedy after a mistake affects a patient or program.

The public terms do not state an output warranty, indemnity or direct correction right against the vendor. Customer support and workflow improvement are part of the relationship, but their obligations following a harmful AI output remain unspecified.

Courier Health centralizes data from hubs, specialty pharmacies, aggregators and other patient access partners. A bidirectional API and alternative data protocols are available, while the shared platform captures documents, status changes and communications. This provides an integration architecture and a clear operational purpose. No named EHR connector, supported EHR version or verified production marketplace connection is attached to the current public offer.

A biopharma patient record is also not equivalent to a provider's clinical chart. The scope of external writeback, identity reconciliation and error recovery remains unspecified for particular systems, even though data exchange is a core platform function.

The offer is an online platform available to office and field users, including access from an iPad or other mobile device. AI processing is described as remaining within the environment, but the current pages do not name a hosting provider, processing region or tenant isolation design. Separate customer hosted or on premises options are not identified. The environment statement is therefore a processing boundary claim without a complete residency map.

Storage, backup, support access and model execution locations remain unspecified. Bringing patient services inside a biopharma organization describes an operating model, rather than proof that the software runs on that organization's infrastructure.

The platform is sold through an enterprise engagement tailored to a therapy program's features, workflows and milestones. Implementation and ongoing account support are part of the described relationship, and the offer can extend an existing hub or support bringing patient services inside the organization. This establishes the commercial posture and the kinds of scope choices involved.

Public subscription rates, billing units, minimum commitments and a price distinction between embedded tools and AI agents are not supplied. The cost of intake volume, additional therapies, data connections and implementation remains unspecified. Customer outcome percentages describe program benefits rather than fees.

The intended users are biopharma patient services, field access and reimbursement teams supporting specialty therapies. The Madrigal case provides a named deployment for Rezdiffra access in MASH, spanning hepatology, gastroenterology and endocrinology workflows. The software coordinates with hubs, specialty pharmacies, providers and internal commercial teams rather than replacing a clinical practice system.

It can be configured for additional therapies and broader portfolios, but comparable results are not supplied for every disease or modality. A named therapy program establishes meaningful setting experience without making the reported improvements universal across the full range of chronic and rare disease programs.