Health Catalyst

Real inference doing real work, inside a company whose own description places expert labor beside the software as a peer rather than a supplement.

The AI is not decorative. Ignite Intelligence combines shared models, software agents, benchmarks and comparison services, and Clinical Cost Intelligence is built as a set of language model agents and statistical and machine learning models that surface cost variation across procedures and service lines, work an analyst would take months to do. Customers such as UnityPoint Health run AI risk stratification for care management on the platform.

Where the grade settles is the surrounding business. The professional services segment exists to configure and operate the technology, and the results the company markets are attributed to that combination. The platform is open to customers bringing their own modeling and business intelligence tools, and Healthcare.AI is sold as products alongside expert services. Strip the models out and a substantial business remains: a healthcare specific data model, standardized measures and registries, identity resolution, an interoperability suite and a large consulting practice.

Autonomy varies sharply across the product line and no single oversight model covers it.

The analytics core is advisory in the ordinary way. Dashboards, registries, benchmarks and costing analysis surface to analysts and improvement teams who decide what to do, and the newer intelligence layer is described as a co pilot, which is the right framing for output that informs a decision made over weeks or months.

The acquired applications behave differently. A care orchestration engine automates steps in a care pathway, and patient engagement and experience products reach patients directly with messages and prompts. Those act rather than advise, and the person affected is a patient rather than an analyst. Software agents in the intelligence layer point the same direction. That is not a criticism of the design, since automated pathway steps and outreach are the point of those products, but it means the company now spans passive analysis and active patient contact under one roof.

What is absent is any unified account of oversight across that range. Nothing published describes which actions execute without a person approving them, what a clinician can override, whether overrides are recorded, or how an orchestration rule is reviewed once live. Oversight of an analytics dashboard and oversight of an engine that messages patients are different problems, and the published material treats the whole product line as one.

The posture is clear for the analytics core and unstated for the products that act.

The approach and the suppliers are named, without the version and update discipline behind them. Clinical Cost Intelligence is described as a multi step, multi agent method: language model agents normalize supply names, map them to an ontology, screen opportunities and draft quality improvement steps, and statistical and machine learning models run the variation analysis. The company names Microsoft Azure and Azure AI Foundry as the AI platform and Databricks as the foundation for model deployment and management. Its value based care page promises open book analytics in which customers can see how every calculation and model works, including risk adjustment, groupers and benchmarks, and customers may point their own tools at the unified data model.

No model version, accuracy, calibration or validation figure is published for any application, and nothing describes retraining or how changes to the underlying language models are handled. For a company that markets transparency, the absence of a single published performance number is the gap.

The model platform and the data platform are named; the full party list and the terms are not. Health Catalyst's April 2025 release says its AI work is built on Microsoft Azure and Azure AI Foundry, and its own webinar material on Clinical Cost Intelligence names Databricks as the foundation for deployment and model management and describes the application as a set of language model agents (supply name normalization, ontology mapping, opportunity screening, quality improvement) working alongside statistical and machine learning variation models. That names the model layer at the level of the platform it runs on.

What holds it below the top is the rest. No subprocessor list is published, the specific language models inside Azure AI Foundry are not named, and nothing states what Microsoft or Databricks retain from clinical and financial content or whether customer data trains any model. Outsourcing is one of the services the company sells, and where outsourced staff sit and what data they reach is not described.

Named deployments with dated results and a stated method, and a headline aggregate that has none. The UnityPoint Health success story describes AI risk stratification over chronic condition, social, admission and readmission data to prioritize patients for care management, and reports that in the 30 months since the program began it reduced spending by more than $32.2 million, with more than 11,000 additional days at home and nearly 2,000 fewer emergency visits, measured by comparing utilization before enrollment with utilization after graduation. The Queen's Health System story reports $22 million in cost savings and $1.9 million in new revenue from patient flow work. As a listed company, customer count and segment revenue are filed, so scale claims can be checked against audited statements.

The method has a limit. A before and after comparison among patients who completed the program has no comparison group, and patients enrolled at a high point of use tend to use less afterward regardless, so the savings figure is an upper bound rather than an estimate of effect. The company's aggregate claim of billions of dollars in measurable results carries no method, period or breakdown, and the company cites no peer reviewed publication.

Solid custodial engineering, a named model platform, and one published phrase that raises the central question without answering it.

The stewardship apparatus is described and credible. Data profiling examines incoming data for quality issues, identity management resolves patient and provider records centrally, and the platform sits under audited security frameworks with named product scope. For a platform whose function is joining records that arrived separately, deliberate identity resolution and continuous quality inspection are the right controls.

The phrase that needs an answer is shared models. Ignite Intelligence is described as combining shared models, agents, benchmarks and comparison services, and benchmarking across a customer base is presented as a benefit. Comparison against peers requires other customers' data in some form, and nothing published says whether that is aggregated statistics computed under contract or model parameters learned from one customer and applied to another.

The language model layer adds a second question. Clinical Cost Intelligence passes clinical and financial content to language model agents running on Microsoft's Azure AI Foundry, and nothing states what is retained, logged or used for training there.

The control posture is documented product by product and the contractual terms remain private.

The strength is scoping. The information security page names which certifications and audited reports cover which products, states framework versions, and gives coverage periods, so a customer buying one application can determine whether that application is in scope. Coverage extends across the data platform, interoperability, population health, patient engagement, clinical quality, patient safety, cost and revenue cycle families. The HISP's DirectTrust accreditation is described as confirming compliance with HIPAA privacy and security requirements for clinical data exchange.

Business associate status is the correct posture for a processor handling identified clinical, claims and financial data on behalf of providers, and the frameworks used map regulatory requirements onto audited control sets.

What is missing is the agreement. No template business associate agreement, breach notification window, liability cap, audit rights statement or data return provision is published. The services segment adds a dimension: consultants and outsourced staff work inside customer environments, and nothing published describes how personnel handling protected health information are governed, screened or bound.

Certifications named by product, framework version and date, with reports a buyer can download without asking. The information security page, linked from the footer, lists SOC 2 Type II reports by product family with their coverage periods, and links public SOC 3 reports for the platform and applications families. HITRUST r2 certifications are listed product by product with dates: the Data Operating System, Ignite and the interoperability platforms (21 August 2024), Embedded applications (12 August 2026), Lumeon (June 2025), Twistle (May 2025) and Upfront (December 2024, with an e1 for a second Upfront product in August 2026). Lumeon holds ISO 27001:2022 for its UK platform, the HISP holds DirectTrust accreditation (July 2026), one product holds TX-RAMP Level 2, and the company carries the KLAS and Censinet Cybersecurity Transparent designation.

Two dates are the buyer's to check. The SOC reports listed for the platform and applications families cover June 2024 to May 2025, so the following period's reports are due, and the core platform's HITRUST r2 reached its two year mark in August 2026 with no newer date shown.

Outside device regulation by function, with the boundary approached from the care orchestration side.

Data aggregation, benchmarking, cost variation analysis, quality measurement and registry reporting are administrative and financial activities rather than diagnosis or treatment, so no clearance is required and none is claimed. Making no regulatory claim where none applies is the honest position.

The company does carry regulatory weight in the reporting domain. Quality measures and registries produced on this platform feed submissions that determine payment and public reporting, and the accuracy of that output is externally consequential even though it is not device regulated. A regulatory and cybersecurity solution line indicates the company treats compliance reporting as a product category in its own right.

The boundary case here is care orchestration rather than a clinician facing assistant. Software that automates steps in a care pathway, deciding when a patient is contacted, when a task is raised and when a protocol advances, sits nearer to clinical workflow than a dashboard does, and how such logic is classified depends on whether it merely executes clinician authored rules or applies inference to a clinical decision. Nothing published addresses which description applies as models are introduced into that layer.

The position is correctly outside device regulation, consequential in payment and public reporting, and unexamined at the one boundary that is moving.

Accountability is assigned and nothing is published on how model behavior is governed or tested. Daniel Samarov is named as Chief AI Officer, and the value based care page promises analytics where customers can see how every calculation and model works. Neither is an evaluation: the home, integration, value based care and security pages carry no bias testing, subgroup results, fairness review, model documentation or statement of what any predictive model takes as its target.

The exposure here is one the research literature has documented. The UnityPoint success story describes risk stratification over chronic condition, social and utilization data to choose who enters care management, which is the task Obermeyer and colleagues showed in Science in 2019 can understate the illness of Black patients when cost stands in for need. Cost variation analysis carries a related risk from the other direction: a model that flags providers or service lines as high cost outliers may be detecting patient complexity and social need rather than inefficiency, and acting on it shifts resources away from the people who generated the cost. Both harms are silent: a patient never enrolled is never told.

Nothing published on what happens when the system is wrong. The home, integration, value based care and security pages carry no indemnity, warranty on analytic output, accuracy commitment, service credit tied to data quality or availability, or route to dispute an output.

The failure modes are financial and specific. A quality measure computed on incompletely ingested data moves a reported score and the payment attached to it. A cost variation finding that reads patient complexity as inefficiency drives a real decision about a service line. A care orchestration rule that misfires reaches patients directly. Each is a loss that traces back to work this vendor performed, and none has a stated remedy.

The services segment makes attribution harder, and against the customer. When the vendor's consultants configure the platform, its advisors design the improvement program and its experts interpret the output, the line between a defective product and a customer that used it badly becomes difficult to draw, and nothing published says how contracts allocate it.

Interoperability is a product line here, built on published standards, with no record system named. The company sells a health information exchange and interoperability platform, acquired as Medicity and now under its own name, and its HISP holds DirectTrust accreditation for secure messaging across record systems. The integration page describes prebuilt connectors for more than 40 healthcare systems, native support for FHIR, HL7 and OMOP, single patient identity across sources, and insights embedded into the tools clinicians use. Named deployments stand behind it, such as UnityPoint Health, which runs its care management analytics on the company's Data Operating System.

None of the 40 connected systems is named in the company's published material, and nothing documents a two way integration with a specific record system that a buyer could verify in a marketplace listing.

Cloud based on named infrastructure, with tenancy, residency and continuity unstated.

The company's April 2025 release says its AI work is built on Microsoft Azure, Ignite is transactable on the Azure Marketplace, and its own material names Databricks as the foundation for model deployment and management. The integration page describes a modular, cloud native infrastructure. Lumeon's certified UK platform indicates at least one product runs outside the United States.

Whether customer data occupies a dedicated or shared tenancy is not stated, and it matters because benchmarking and comparison services need some cross customer computation, so the tenancy boundary and the analytical boundary are not obviously the same line. No region or residency position, recovery objective, availability commitment or failover description is published, for a platform health systems run quality reporting and cost management on continuously. Whether the acquired applications share one deployment is not described.

No price is published, and how the product is sold is discoverable from the company's own filings and releases. Health Catalyst reports two segments, technology (sold as cloud subscriptions and term licenses with maintenance and support) and professional services (analytics, implementation, advisory, outsourcing and improvement services), which shows that cost follows the modules licensed and the services attached. Ignite and Healthcare.AI Enhanced Analytics are transactable on the Azure Marketplace, which lets an organization buy against an existing Microsoft commitment.

What cannot be sized is any figure. No unit of charge, range, tier, implementation fee or minimum commitment is published on the home, integration, value based care or security pages, and the ratio of services to technology in a typical first year, the most useful number a buyer could have here, is not given in a form a buyer can apply to itself.

Broad coverage across function, setting and buyer type.

Functional coverage runs across clinical quality and patient safety, cost and revenue cycle, population health and value based care performance, patient engagement, care orchestration, regulatory reporting and health information exchange. Setting coverage extends from acute care into ambulatory, with a product line for acquired and independent ambulatory practices where access, panel management, provider productivity and referral management are the operative problems. Health information exchange participation extends reach past a single organization to regional data sharing, and Lumeon's UK platform adds a market outside the United States.

The supporting assets are the reason this is coverage rather than a list. Standardized measures and registries mean specialty measurement is prebuilt, and prebuilt connectors for more than 40 systems mean a new setting is an onboarding task rather than an engineering project. Scale is stated at more than 1,100 organizations, alongside audited filings.

Depth per area is the gap. Breadth of this kind is assembled partly through acquisition, and nothing published establishes how deeply any single specialty is served relative to a focused competitor, or how well the acquired products integrate with one another rather than being sold together.