Cedar Gate Technologies

Predictive modeling is claimed across the analytics layer and one model is named; the assets that make the company hard to replace are the payment engines and the data foundation underneath them.

The analytics offer includes prediction of health and utilization risks, and a 2023 release describes a predictive AI model scoring each woman's probability of getting a recommended mammogram, applied to 2.4 million women in the benchmark database. That is a learned component doing real work in outreach targeting.

Remove the models and most of what is sold remains. A data lake that ingests, cleanses, normalizes and stitches more than 50 data types still has value. The bundles engine converting fee for service claims into a bundled claim, and the capitation engine administering agreements against changing eligibility, still work, because adjudication is deterministic by design. AI is a feature layer on a product whose value stands without it.

The platform executes financial transactions automatically, so the oversight that matters concerns money rather than patients.

The analytics and care layers behave conventionally. Insights surface to plan and provider staff, care management supplies workflow to people who then act, and nothing touches a patient without a person.

The payment layer is different in kind. The bundles engine automatically converts fee for service claims into a single bundled claim and facilitates and manages payment between payers and providers, and the capitation engine administers agreements against changing eligibility. That is software adjudicating and settling money without a human approving each transaction, at a volume the company describes in the millions of bundles. It is also the correct design, because adjudication at that volume cannot be manual and determinism is precisely what makes it trustworthy.

The oversight questions are therefore financial rather than clinical, and the published material answers none of them. What happens when an episode is adjudicated incorrectly, who detects it, whether reconciliation is automated or manual, what the dispute path is for a provider who believes a bundle was settled wrongly, and whether adjudication logic changes are versioned and communicated before they take effect are all unaddressed.

An error here does not harm a patient. It underpays or overpays a provider, potentially systematically and potentially for months before anyone notices, and the party least able to detect it is usually the one being underpaid.

The architecture is described clearly and the models are described by function only. The data lake, the application layers above it and the payment engines are explained well enough for a technical evaluator to follow how data moves, including automatic conversion of fee for service claims into a single bundled claim. On the model side, the platform is said to predict health and utilization risks with predictive and prescriptive analytics, and one model is named by what it does: an October 2023 release describes a predictive AI model that scores the probability a woman will get a recommended mammogram, with a threshold below which she is classed as unlikely.

No architecture, training data, accuracy, calibration or validation is published for that model or any other, and no retraining or versioning practice is described. How the benchmark database is built and how peer groups are matched is not published either, which makes a benchmark result hard to interpret and impossible to challenge.

The architecture is described, the host is named, and no model provider is named. Amazon Web Services hosts the data lake, which draws more than 50 data types from a library of more than 350 source vendors, cleanses and enriches them and feeds the analytics, care and payment modules. No subprocessor list, data licensor, model provider or component inventory is published on the home, data or information security pages.

The benchmark database is a data supply chain of its own. The company describes more than 12 million member lives in a proprietary national benchmark database and uses it for its own published analyses, such as a 2023 study of 2.4 million women; where those lives came from and under what terms they may be used is not stated.

The labor chain is named in outline: more than 700 employees in Nepal, with nothing on what data that team reaches or under what controls. Ownership now sits with IQVIA, and nothing describes how component changes under the new owner reach customers.

A long list of named customer results and headline figures large enough to require the method that is not supplied.

The better evidence is the specific and modest kind. A cardiovascular care network reports a 26 percent reduction in spending per cardiac patient after moving to a bundled arrangement built on the adjudication software. A medical university with more than 100 locations reports a 3 percent improvement in comorbid chronic kidney disease patients avoiding hospitalization, with average savings of $36,100 for those who did. Both are plausible in magnitude, both are attached to identifiable organizations, and the second states its denominator population.

The headline figures are a different matter. A bundled payment case study presented at an industry summit in 2026, alongside two named partner organizations, reports 97 percent patient satisfaction, 91 percent avoided inpatient hospital care, a 60 percent reduction in total cost of care and a 50 percent reduction in surgical rates. Those numbers are extraordinary. A 50 percent reduction in surgical rates and a 60 percent reduction in total cost of care would each be a landmark result in health services research, and presented together without a denominator, a time period, a comparison group or a stated method they cannot be assessed at all. The most likely explanation is a narrow selected cohort in a single program, which is a legitimate result described in a way that invites a much larger reading.

The home, data and information security pages cite no peer reviewed publication and no external validation of the data or measure logic.

The data engineering is described concretely, and the benchmark database raises a secondary use question the published material does not close.

The custodial description is concrete. Data is ingested from disparate sources, cleansed, normalized into standardized files and stitched directly to members, with automated enrichment producing usable information. Member stitching is the hard part of aggregation and naming it as a distinct step indicates the problem is treated seriously rather than assumed away.

The question is the benchmark asset. Buyers are offered comparison against an anonymized database described as covering 12 million member lives, which is presented as a benefit and is a real one, since a cost and utilization figure means little without a peer group. What is not described is where those 12 million lives came from, and the plain reading is that they came from customers. If so, nothing published says what contractual permission allows one customer's data to inform another customer's benchmark, what anonymization standard was applied and by what method, whether the process was externally reviewed, or whether a customer may decline to contribute while still consuming.

The word anonymized is doing considerable work in that description and is not defined anywhere in the published material. Claims and utilization data at member level is among the more re identifiable data that exists, and the difference between a statistical aggregate and a de identified record set is the difference between a low risk and a meaningful one.

Model data governance is unaddressed, with no statement on whether customer data trains models or whether learning crosses customer boundaries.

Status is stated and supported by a substantive privacy document; the agreement is not published. The information security page says Cedar Gate complies with HIPAA's regulatory requirements, and the enterprise data management page describes the technical safeguards applied to protected health information on Amazon Web Services: secure transmission, authentication, access controls, integrity controls and auditing of extracts sent to third parties. The footer's privacy policy is IQVIA's corporate policy, which covers personal information processed in the performance of IQVIA's services; it does not mention HIPAA or business associate terms.

No business associate agreement, breach notification window or subcontractor position is published. The workforce question stands: the company's history page puts more than 700 of its more than 1,000 employees in Nepal, and nothing says what protected health information staff outside the United States can reach or under what controls.

Two recognized attestations are named as the company's own, without the artifacts. The information security page, linked from the footer, states a current SOC 2 Type II report and HITRUST risk based, two year (r2) certification, and names the platforms the HITRUST certification covers: Analytics, Bundles Adjudication, Care Management, Revenue Cycle Management and User Management. It also states HIPAA compliance and a security program based on the NIST Cybersecurity Framework.

Three questions remain for the buyer. No certification date or SOC 2 reporting period is given, and an r2 certification runs for two years. The HITRUST scope list does not name Capitation Adjudication or Enterprise Data Management, the data lake every module draws on. And no report can be opened without asking, since there is no trust portal.

Outside device regulation for most of what it sells, and regulated in the financial and program direction instead.

Analytics, benchmarking, care workflow and payment adjudication are administrative and financial functions rather than diagnosis or treatment, and no clearance is claimed. The platform menu also lists clinical decision support and utilization management modules, but nothing published says what they recommend or to whom, or whether either issues clinical recommendations and under what regulatory position.

The regulatory weight sits elsewhere, and it is real. Software that adjudicates claims and settles payment between payers and providers operates inside a body of state insurance regulation, prompt payment requirements and contractual arrangements with defined timelines, and a capitation engine administering eligibility touches enrollment rules. Support for a mandated federal episode payment model introduced in March 2026 places part of the product inside a program where participation is compulsory for affected hospitals and where the rules are set by a regulator rather than negotiated.

What is not described is how the company demonstrates conformance in any of those directions. Nothing published covers how adjudication logic is validated against program rules, how rule changes are absorbed when a federal model updates its specifications, or what assurance a customer receives that a settlement complies with the arrangement it was meant to implement.

That is a narrower question than device clearance and a more immediate one, because a mandated program has deadlines and a vendor that lags a specification change puts its customer out of compliance rather than merely out of date.

Nothing published on how model behavior is governed or tested. No bias testing, subgroup results, fairness review or model documentation appears on the site. IQVIA, the owner, published five responsible AI principles in an August 2024 blog post; it predates the acquisition, speaks for IQVIA's own AI rather than anything Cedar Gate adopts for its products, and names no process.

The mechanisms at risk are specific. Cohorts built by risk profile from claims and utilization history, and the mammogram model that classes women as unlikely to screen, sit inside the failure Obermeyer and colleagues documented in Science in 2019, where past spending used as a proxy for need understated the illness of Black patients. Bundled payment adds a second: if episode pricing under adjusts for social complexity or comorbidity, providers serving harder populations lose money on bundles, and the analytics that show which episodes are profitable also show which patients to avoid. Benchmarks against the 12 million life database add a third, reading an organization serving a deprived population as an underperformer if peer groups are not adjusted.

Nothing published on what happens when the system is wrong. The home, data and information security pages carry no warranty on analytic output, accuracy commitment, indemnity, service credits or dispute process, and the home page's "we succeed when you succeed" names no measure and puts no fee at risk.

The distinctive exposure is adjudication. When software converts fee for service claims into a bundled claim and settles payment between a payer and a provider, an error is money moved incorrectly, and systematic error in bundle construction or episode attribution could underpay providers across every episode processed. Capitation against changing eligibility has the same shape. A provider underpaid through the vendor's adjudication holds its contract with the payer, not with the vendor, so the dispute runs through a party that did not perform the calculation.

Integration is claimed broadly and no system is named. The enterprise data management page describes ingesting more than 50 data types, claims, clinical, eligibility and benefits among them, from a library of more than 350 source vendors with a defined layout and load specification for each, then cleansing, normalizing and stitching records to members. The payment modules are said to connect with claims engines for adjudication and payment.

None of those sources or claims systems is named, no interface standard is stated, and nothing says whether findings write back into a clinical record or care workflow. A buyer cannot verify a single connection from the published material.

A single hosted option on a named cloud, with location implied rather than committed. The enterprise data management page says the platform runs on Amazon Web Services and describes the data model it uses there. No region, residency commitment, tenancy model, recovery objective or availability commitment is published on the home, data or information security pages.

Two features make the gaps matter. The payment engines adjudicate claims and settle money between payers and providers on schedules providers depend on, so availability during a settlement cycle is a payment question rather than a convenience. And the company's history page puts more than 700 of more than 1,000 employees in Nepal, which says nothing about where data rests and raises where it is accessed from.

The shape of the purchase is discoverable; the price is not. The platform is sold as composable modules on one data lake, covering data management, benefits and value based care analytics, care management, clinical decision support, utilization management, and bundles and capitation adjudication, alongside two service lines, actuarial consulting and bundles program management. The published material makes clear that cost follows the modules taken and whether services are added.

What cannot be sized is any figure. No unit of charge, range, tier, implementation fee or minimum commitment is published, and adjudication, which is transactional work, has no stated per bundle, per claim or per member rate.

Coverage of payment models is the distinctive breadth here, and it is more unusual than coverage of clinical conditions.

The platform addresses primary care attribution, shared savings, prospective and retrospective bundles, and capitation, across commercial, Medicare and Medicaid lines, with support for a mandated federal episode model launched in March 2026. Bundles and capitation are harder than shared savings because both require adjudicating payment rather than reporting on it.

Buyer coverage extends past health plans and provider organizations. Health plans, provider organizations, self funded employers, third party administrators and adviser partners are all addressed, and the employer and administrator segment is served with its own benefits analytics product. Self funded employers are a different kind of customer, with different data, different questions and no clinical operation of their own, and serving them is not a variation on serving a health system.

Clinical coverage follows from the payment models rather than from condition specific content, with bundles built around procedural episodes such as cardiac and musculoskeletal care and analytics addressing chronic disease populations.

The evidence for depth is thinner. Breadth assembled through four acquisitions raises an unanswered question about how completely the parts function as one product, and no installed base figure by segment is published.