Fourier Health

The models are the product. Fourier Intake classifies, links, retrieves and structures every inbound page (faxes, PDFs, scans, HL7, FHIR and exchange records), Fourier Insights assembles specialty summaries from that output, and Fourier Infrastructure sells the structured data through an API.

The company's engineering post describes the pipeline: two-stage page classification, language model patient matching, hybrid retrieval with reranking and model filtering, and an agent that assembles schema-bound outputs, with rules used only for urgency flags. Human reviewers check a sample of outputs rather than producing them. Remove the models and there is nothing left to sell.

The marketing promises that every result is reviewed. The Intake page says "Every result reviewed by healthcare experts, so you get safe, reliable data, every time", and the home page says clinicians ensure every output is reviewed.

The company's own engineering post describes something narrower and better specified: human QA on a sample, weighted toward new specialties, new customers and components with low QA scores, an AI-assisted layer that flags likely errors first, and risky patient-record merges routed to a human queue before processing. A team that believes every result is reviewed and always right will not build the check that sampled review leaves. Every extracted field is cited to its place in the source document.

The approach and the model options are named without version or update practice. The engineering post (April 2026) describes each stage: OCR and two-stage page classification over 100-page chunks, language model patient matching with a risk check, hybrid semantic and lexical retrieval with reranking and a model filter, bounding box reconstruction for citations, rules for urgency, and an agent that fills specialty schemas; the home page lists Claude, GPT, Med-PaLM, in-house and custom models as supported options. Which model runs a given customer's pipeline, its version, and how customers learn of changes are not published.

Model providers are named as options and the rest of the chain is not. The home page lists Claude, GPT and Med-PaLM alongside in-house and custom models, and customers can bring their own. QA reviewers outside the customer, unnamed service providers (hosting, data analysis) and the partners who receive de-identified data are also in the chain. No subprocessor list, reviewer contracting terms or retention terms for data sent to a model provider are published.

Named customers and vendor-reported results. The home page shows Mayo Clinic, One Oncology, Tennessee Oncology, Summus, Telemedica and Inception among its customers, and reports an 83.9 percent reduction in records reviewed per patient and 82 percent clinician satisfaction.

Five case studies describe unnamed clients with figures: a 34,000 patient academic system ($655K net annual savings), an academic referral center (28 to 17 days to first appointment), a community oncology network (1.8 million pages a year, 41,000 hours saved), a fertility network (six hour median turnaround) and a cardiovascular value-based care organization (4.2 times return, $2.25M shared savings). The accuracy benchmarks are the company's own. No independent or peer reviewed evaluation is published.

Customer data is reused. The company's engineering post says training data, evaluation sets and retrieval configurations are shared across customers in the same specialty; the privacy policy (15 September 2025) says de-identified data is used to refine its algorithms and machine learning and may be shared with partners and in benchmarking; and Fourier Infrastructure sells clinician-labeled, de-identified datasets to AI and research partners.

The privacy policy states a general retention of up to seven years and says health data is stored and processed on HIPAA-compliant services. Clinician reviewers outside the customer read records during QA. No safety engineering for generated summaries is published beyond sampled review.

A privacy document reaches the service, and no business associate status is stated. The privacy policy covers creating summaries from medical history and says the company uses HIPAA-compliant services for health data, and the home page shows a HIPAA compliant badge. No business associate agreement or its terms are published. QA reviewers outside the customer read records, and de-identified data flows to partners, so the agreement's scope matters.

A recognized certification is shown as the company's own, without the artifact. The home page carries an "AICPA SOC 2 Type II certified" badge; the report period and scope are not stated, and no trust center or report request path is published. The intake page says records are encrypted in transit and at rest with a complete audit log. QA reviewers reach records from outside the customer, and nothing says whether the SOC 2 scope covers that reviewer access.

No device claim is made and no regulatory position is published. FDA's 510(k) database lists no clearance for the company. The pipeline flags urgency with rules on lab thresholds for coded findings, extracts diagnoses, and lists decision support and care gaps among its outputs; every output is cited to source for clinician review.

A named quality process, no subgroup results. The company grades outputs against a rubric through human QA, directs sampling toward new specialties, customers and weak components, and uses an AI-assisted QA layer trained on reviewer feedback; its engineering post publishes per-label and retrieval metrics against frontier models. No results by age, sex, language or record source are published.

The company also says training data, evaluation sets and retrieval configurations are shared across customers in the same specialty, and its case studies report revenue and risk capture alongside clinical results.

Error rates are published with a method and, for one figure, a denominator. The approach page reports F1 of 0.96 on medical data classification against expert-adjudicated ground truth, and 3.5 times the accuracy of manual intake on an internal evaluation of 2.4 million documents; the same page charts an information miss rate of 0.2 percent against named frontier models; the engineering post reports per-label F1 of 0.94 to 0.98, page-level recall of 0.995 and token-level recall of 0.998.

These are the company's own evaluations, not independent ones. Every field is cited to its source with bounding boxes a clinician can check. No warranty, indemnity or remediation commitment is published.

Named systems with writing described, and nothing to verify. The company says it embeds in Epic, Cerner, Athena, Meditech and others and feeds structured data directly into the EHR, and ingests HL7 v2 messages, FHIR bundles and CCDs pulled from health information exchanges; output is returned as FHIR JSON through its API. No marketplace listing or integration documentation is published, and no case study names the EHR it runs in.

A single hosted service with the location implied. Records arrive by fax, SFTP, API or upload and results return to the EHR, inbox, API or web app; the company says customers can bring their own model, use its models, or combine them. The privacy policy says health data is stored and processed on HIPAA-compliant services without naming them. No provider, region or residency commitment is published, and QA reviewers' locations are not stated.

No price is published, and the products are discoverable. The company sells Fourier Intake, Fourier Insights and Fourier Infrastructure (API and datasets), and its case studies express value as savings, payback (3.5 months at one academic system) and return on investment. No unit of charge, tier or implementation fee is published, and nothing says what the clinician QA layer costs.

Named specialties with case studies behind part of them. The company builds specialty schemas (cardiology among them) and reports work in academic health systems, community oncology, fertility screening and cardiovascular value-based care, for health systems, specialty networks and AI research partners. The case studies are vendor published and unnamed. No age, population or specialty exclusions are published.