Mandolin

Mandolin's agents interpret referral documents, compare charts with payer policies, navigate portals and make phone calls to complete specialty drug administration. Learned language, vision and multimodal reasoning are central to that work, rather than an optional reporting feature. The same operational context follows a patient from intake through benefits verification, authorization and claims follow up.

Deterministic checks can handle defined fields, while ambiguous clinical documentation is prepared for expert review. Without the models, the current offer would lose the ability to interpret varied records and carry work across systems, which is the principal service being sold.

Agents can enter records, make payer calls and submit authorization requests, so their role extends beyond drafting suggestions. The current design philosophy separates deterministic work from ambiguous material and keeps treatment necessity and safety decisions with a clinician. Its examples route contradictory diagnosis information toward human review, cross check inconsistent measurements and preserve source wording behind classifications.

Call transcripts and action logs support subsequent review. The public workflow description does not specify confidence thresholds, mandatory stop conditions or an override process for each task. Broad automation language therefore coexists with a meaningful but incompletely specified human oversight structure.

The technical account describes a shared operational context that combines vision, language and multimodal reasoning across structured records, clinical notes and other documents. Agents interact with EHR interfaces, payer portals and telephone systems, while defined checks can be handled through rules. This gives a functional architecture for how work is coordinated. It does not identify a foundation model provider, model family, version or update schedule.

Training data are discussed primarily through privacy rules for deidentified material, rather than through a technical description of datasets and evaluation. The current material explains behavior more fully than the underlying model stack.

The privacy policy contemplates third party AI developers and model development platforms, with deidentification or BAA based safeguards for shared information. It also describes categories of service providers such as hosting, analytics, support and security. Specific model providers and a named subprocessor roster are not included.

As a result, the public account does not establish which organizations process a clinical document or call transcript, nor the retention and training terms that apply at each stage. A general promise to bind PHI subcontractors is meaningful, but it does not enumerate the model processing chain for the current agents.

The homepage describes a document intake workflow falling from twenty minutes to three minutes per document, an eliminated prescription backlog and a separate infusion operation serving more than 4,500 patients monthly. These accounts identify customer types and operational baselines, but not customer names, measurement dates or a controlled comparison. Greylock names Vivo, FlexCare, OI Infusion and TwelveStone as deployments in its June 2025 investment announcement.

Those names substantiate use, while the scale figures describe reach rather than benefit. The available accounts do not connect a named customer to a dated, reproducible performance study or independently measured patient outcome.

Customer records and AI generated insights can contain PHI. The policy permits deidentified or aggregated information to train, validate and improve models, and requires appropriate safeguards or explicit BAA permission for outside model development. Retention follows customer instructions, contractual requirements and the purpose of processing, followed by deletion or anonymization when information is no longer needed. Encryption, access controls and incident response are also stated.

There is no numerical retention schedule for prompts, transcripts, outputs or backups, and no detailed deidentification method. The policy therefore supplies categorical commitments without a complete operational account of the AI data lifecycle.

The service privacy policy identifies Coral Automation, doing business as Mandolin, as a business associate when providing services to healthcare organizations. It states that PHI handling is governed by an executed BAA and that subcontractors handling PHI must also sign one. The policy covers the healthcare platform itself, including records and AI outputs, rather than website tracking alone. A complete public BAA and a product or subscription level availability schedule are not included.

Patients are directed to their healthcare organization for record rights, with Mandolin assisting under the agreement. This establishes a substantive service posture while leaving contract specifics private.

The homepage displays a HIPAA compliance badge, and the privacy policy describes encryption, access controls, vulnerability management, staff training and incident response. These are security practices and compliance assertions, rather than evidence of an independent security assessment. No named SOC 2 report, ISO certificate or comparable third party artifact is attached to the current public offer. Regular security assessments are mentioned without an assessor, scope or report.

The badge therefore does not establish outside certification. The product's access to EHR accounts, payer portals and patient records makes the assessment boundary especially relevant to understanding any future assurance claim.

The described purpose is administrative support for access to specialty therapies: collecting records, checking benefits, assembling authorization submissions and following claims. Medical policy review involves clinical information, but the current design explicitly reserves decisions about whether treatment is necessary and safe for a clinician. No FDA authorization is attached to this administrative product.

Examples involving laboratory values, diagnosis codes and drug histories describe the information that staff must verify, rather than an independently cleared diagnostic function. The boundary is particularly relevant where data entry affects later treatment decisions, since administrative positioning does not make an incorrect record harmless.

The privacy policy commits to mitigating bias through continuing evaluation and improvement. Product design articles add more specific review principles: preserve clinical context, flag conflicting information and keep consequential treatment decisions with clinicians. These are useful commitments, but they do not include subgroup evaluation methods or results for patient populations, languages, drug categories or payer workflows. No independent AI governance assessment is identified.

The available detail also leaves the process for approving model changes, tracking regression and resolving a recurring bias issue unspecified. Action logs can support investigation of an incident without establishing that systematic disparities have been measured.

Benefits verification combines payer call execution with transcripts and audit trails. That record can help staff reconstruct what was asked, what the payer answered and how an authorization relates to a later denial. The current denial article explicitly acknowledges that transcripts and authorization do not guarantee recovery, and that exceptional cases still need human coordination. These features support traceability and correction, not a contractual transfer of reimbursement risk.

There is no stated output warranty, indemnity for a mistaken verification or defined financial remedy for an incorrect submission. The financial consequences of relying on agent output remain governed by private commercial terms.

Mandolin reads referral forms, laboratory reports and clinical notes and enters the extracted information into an EHR. Benefits and claims work can use payer portals, fax and telephone, so the product is designed to operate through existing interfaces without requiring an API integration. That is a concrete interaction pattern, but no named EHR connector, supported version matrix or verified production marketplace listing accompanies it.

The available detail does not establish the limits of record writeback, reconciliation or failure recovery. Interface automation can connect a workflow operationally while still leaving the reliability of each specific system combination unresolved.

The privacy policy places the service's servers and central database in the United States and describes possible transfers there for users outside the country. The offer is a hosted healthcare platform; separate on premises or customer cloud options are not identified. A cloud provider, tenant isolation design and region specific processing commitments are also unspecified.

The US location is useful context, but it does not map every path taken by model inputs, generated outputs, support access and backups. Encryption and contractual safeguards address protection during processing without supplying a complete residency and deployment specification.

Sales contact is required to establish the commercial terms. The current offer spans intake, benefits investigation, patient cost estimation, prior authorization and claim follow up, but there are no public subscription figures, transaction rates or package boundaries. It is unclear whether charges depend on locations, patients, workflows, completed tasks or another unit.

Customer stories describe staffing changes and faster processing, which indicate the intended economic benefit without revealing fees. Funding amounts also concern the company rather than the cost of its software. Implementation charges, minimum commitments and the treatment of failed or exceptional tasks remain unspecified.

The focus is specialty therapy operations at infusion centers, specialty pharmacies and health systems. Greylock's June 2025 account names Vivo, FlexCare, OI Infusion and TwelveStone among live users, providing identifiable examples of the intended setting. Workflows span referral intake, benefits, medical policy review and reimbursement rather than a single disease.

Current clinical operations articles discuss the documentation needs of multiple sclerosis and thyroid eye disease therapies as illustrations. These examples explain the complexity of the work, but they are not specialty specific validation studies. Performance across drug categories, payer policies and different delivery settings is not established uniformly.