Particle Health

The network is the product and the AI is a layer on it. What a customer buys first is record retrieval: connections to national and state exchanges, record location and patient matching at a reported 320 million patient scale. The AI sits in Snapshot's summarization of retrieved records, in the parsing, de-duplication and normalization of medications, labs, procedures and conditions, and in Signal, Workbench and Navigator's alerts and care journey views. Remove the models and the retrieval network, its API and its ADT and pharmacy feeds would remain the thing sold.

The operation everything depends on is sold as infallible. The insights platform page says its algorithms link patients to a universal identifier, "guaranteeing that all data retrieved is matched to the intended individual". Patient matching across organizations is a probabilistic judgment on names, dates of birth and addresses, and the company's own engineering writing argues that alerts fail at identity rather than delivery.

The guarantee invites exactly the wrong behavior: a buyer who believes every record is matched to the right person will not check, and a record wrongly attached to a patient looks exactly like a correct one. The other oversight the company describes is on who may query: customers go through onboarding, must follow the exchange frameworks' rules and are removed if they do not, and audit APIs log every query.

The approach is explained in engineering terms, without model detail. The company's blog explains why conformance to C-CDA standards is not comprehension and why template-based parsing drops clinically important medications, argues that discharge alerts fail at patient identity rather than delivery, and describes building a discharge summary that does not wait for the hospital's.

The insights page describes the record locator's search logic and a proprietary tabular data format beyond C-CDA and FHIR. No model, version, benchmark or accuracy figure is published for Snapshot's summarization or the parsing layer.

The host is named and no model provider is. The security page says Google Cloud provides authentication and authorization for all inter-service communications, placing the platform there. Snapshot's AI-powered summarization and the parsing and enrichment layer are described without saying whether any outside model provider processes record text, and no subprocessor list is published; the privacy policy's service provider language covers the website. The upstream parties are named instead: CommonWell, Carequality, eHealth Exchange, state HIEs, ADT feeds and pharmacy networks.

Named customers and vendor reported reach, with no retrieval quality measures. The home page quotes One Medical on gathering external health data before a first visit, xCures on growth and scalability, and Cureatr on pharmacist recommendations from discharge summaries, and the company reports 320 million patients, more than 70,000 organizations and 90 percent of patients found on average. No clinical evidence applies, since the product retrieves rather than decides.

What would matter is operational: what share of queries return records, how completeness varies by region and record system, latency, and medication capture completeness. None is published, and a national average hides wide regional variation.

One categorical commitment is published, without a retention schedule. The security page states that all protected health information is de-identified when used for internal analytics, and that data at rest is encrypted. Nothing states whether retrieved records are kept after a query is fulfilled or only passed through, how long, whether de-identified records train the Snapshot summarization or other models, or what de-identification standard is applied.

The privacy policy covers website visitors only. The records concerned come from organizations across the country that have no relationship with Particle, contributed to exchange networks for treatment and operations.

HIPAA compliance is claimed and no service document is published behind it. The security page and insights page describe a HIPAA compliant platform, and the security page says it exceeds all legislative requirements for access to protected health information. The privacy policy (effective 29 July 2026) applies solely to website visitors, and no business associate agreement or subprocessor disclosure is published.

Permitted purpose is the operative control: in exchange networks the querying party asserts why it wants a record, and the intermediary's checks stand between that assertion and the record. In 2024 a major record system vendor complained that some Particle customers queried under treatment purpose for other uses; the framework's published resolution found Particle had carried out due diligence at onboarding that did not surface the customers' misstatements, and those customers were removed.

The company presents two attestations as its own. Its security page lists a HITRUST certification and SOC 2 compliance for the API platform, and the insights platform page describes a platform that is SOC 2 and HITRUST certified. The page also describes OAuth 2.0, multi-factor authentication, SAML single sign-on, audit APIs that return login attempts, full query paths and notification history, AES encryption at rest and mutual TLS, with inter-service authentication on Google Cloud. No certification period, SOC 2 type, scope, report or trust center is offered, though HITRUST status can be checked against HITRUST's public registry.

Data exchange infrastructure is the core product, and it carries no device claim. The Med Rec product comes closest to care, promising de-duplicated medication data to help ensure the right patient gets the right medication without contraindicated prescribing, but it supplies records rather than recommendations. Answering to the rules that govern exchange is not a regulatory pathway.

Those rules are active: information blocking provisions, the participation agreements of the national frameworks with their own dispute machinery, and competition law, through the company's continuing antitrust action against a major record system vendor over access decisions.

No governance framework or fairness statement is published, and the question that matters is patient matching. Deciding that records held by different organizations belong to one person relies on name, date of birth and address, and degrades where names are transliterated or hyphenated, addresses change often, or records sit under a former name; those conditions correlate with population groups, so missed matches are not randomly distributed.

A missed match shows a clinician an incomplete history with nothing to reveal the gap. The company's engineering writing on identity failures in alerting is adjacent analysis that shows it understands the problem, but no match rate by group is published.

Mechanisms exist to check an output, and nothing published stands behind it. Audit APIs return every query path and notification event, and the insights platform keeps provenance records for retrieved data, so a customer can trace where a record came from. The company's engineering writing names its own failure modes, that template based C-CDA parsing drops medications and that alerts fail on identity, which marks where errors are most likely to surface.

No warranty, accuracy commitment or remediation is published, and the matching guarantee on the insights page has no error rate, method or denominator behind it. The person harmed by a false match is a patient whose record is attached to someone else, with no route to the intermediary.

This is the product, with named networks and named customers behind it. A single bidirectional RESTful API returns FHIR R4 or C-CDA data from CommonWell, Carequality, eHealth Exchange and state HIEs, plus ADT feeds and pharmacy fill data, reaching what the company reports as more than 70,000 organizations and 320 million patients and finding on average 90 percent of patients nationally.

Its record locator narrows searches by geography, known conditions and recent activity, and Particle for Platforms handles user creation, usage policies and multi-network administration for health IT vendors. Named customers include One Medical, xCures and Cureatr, and documentation and a free sandbox are public. The residual is currency: access through one framework was restricted during a 2024 dispute with a major record system vendor.

A single hosted API on Google Cloud, with no region or retention stated. The security page places inter-service authentication on Google Cloud, and delivery is an interface consumed by customer applications or the platform's own user interface. What matters most for an intermediary is what it keeps: whether retrieved records pass through to the customer or persist to serve later requests and improve matching. Nothing published says which, and during the 2024 dispute the complaining vendor asked Particle to delete data obtained through the framework, so responding organizations care.

No price is published. Developers can sign up to a free sandbox through the developer portal, and the products (core APIs, FOCUS curated datasets, specialty search, ADT and pharmacy feeds, Snapshot, Signal, Workbench and Navigator) are otherwise sold by demo, with a platform arrangement for health IT vendors serving many clients. The pricing unit is not stated.

Per query, per patient record retrieved, per alert and subscription pricing behave very differently when value scales with retrieval volume, and a per query model can make broad panel searches expensive. FOCUS is pitched as cutting cost by returning a curated subset of attributes, which suggests volume of data retrieved drives price.

National reach, read on two dimensions. Retrieval covers wherever participating organizations hold records, across primary care (on-site, virtual, home health and value-based care), specialty care with condition-specific data for oncology, endocrinology, cardiology and more, and platforms and record system vendors serving many downstream clients. Buyers named are value-based care organizations, payers, health systems and developers.

The company also sells the same data to value based care organizations and payers for care gap closure and HEDIS, MIPS and Star Ratings reporting, with Signal alerts sold for readmission reduction. Snapshot, sold on its own product page, uses large language models to turn retrieved records, clinical notes included, into point of care summaries for providers.

Permitted use is narrower than reach: it is set by the purpose a customer asserts and the framework carrying the query, and a payer querying sits in more contested territory than a provider querying for treatment.