Anterior

The AI system is the product. Florence, built on frontier models, reads charts, faxes, structured feeds and policy text and produces the review outcomes, evidence packets and recommendations that the four programs (Triage, Flow, Relay and Steward) package for health plans. Remove the models and what remains is an ingestion layer, a decision ledger and the company's implementation staff, with nothing to sell: there is no record system, claims platform or outsourced review service underneath. The clinicians the company embeds with each plan configure and tune the system, which is delivery rather than a separate product, and the review organizations that run Florence, such as MedWatch and HealthHelp, are customers and partners rather than part of the offer.

Which parts of the reasoning are the company's own is not clear: the current material describes the system as built on frontier models, while earlier material positioned it against general purpose models.

A described oversight structure, and a correctness promise in the company's own copy that tells buyers it is unnecessary. Florence completes work automatically only where criteria are clearly met and otherwise escalates to a qualified human with a sourced evidence packet; it runs as background agents or under human in the loop review; and the July 2026 Stellarus partnership announcement states that the AI may recommend approval or route a case for review but will never deny care, with all final determinations made by licensed clinicians. That is a boundary published both ways, with no threshold for clearly met.

The home page sells payment integrity as "Claims payments accurate every time". That promises the output is right, which covers what a review misses as well as what it catches, so it grades here whatever the escalation design, and the company's own 97 percent clinical accuracy figure for MedWatch, across utilization management, care management and payment integrity work, says it is not every time. The copy also points two ways: the home page offers automated approvals and a 2025 case study describes instant approval of high confidence requests, while the announcement says licensed clinicians make every final determination.

Frontier models are the stated base, and none is named. The Florence page says the system is built on frontier models that reason across charts, faxes, structured feeds and policy text, uses FHIR internally, and checks its own work with language models in production; the trust center's summary calls it a proprietary architecture designed to replicate clinical reasoning. Earlier material positioned the product against general purpose generative models and the current page builds it on them, so which parts are proprietary is unclear.

What is missing is which models, what data any adaptation is trained on, and a version and update practice that tells a plan when the system changed. The version control the company publishes applies to each plan's encoded policy standards, not to the models. The 99.24 percent clinical accuracy figure, which the company's February and July 2026 announcements attribute to validation by KLAS Research, and the 97 percent MedWatch figure carry no published method.

The architecture is described and no model provider is named. The Florence page says the system is built on frontier models that reason across charts, faxes, structured feeds and policy text, with checks run by language models in production, and a 2025 Points of Light case study on the site of the utilization management partner HealthHelp describes Florence as a clinical reasoning engine powered by a large language model. Which models, from which providers, is not stated on the home, Florence or privacy pages. The privacy policy lists service provider categories including hosting and artificial intelligence without naming any, and it covers the website and application rather than the protected health information processed for plans. The trust center may hold a subprocessor list.

The hosting provider and region are not named either, and nothing states the retention and training terms each provider is bound to once a plan's clinical records reach it.

A named customer and vendor reported figures, with the independent validation the company cites not published. The home and Florence pages report that MedWatch, a national utilization management organization, achieved three times clinician productivity across utilization management, care management and payment integrity work, with 92 percent satisfaction and 97 percent clinical accuracy. The February 2026 funding announcement cites 99.24 percent clinical accuracy in live production, which the company says KLAS Research independently validated, and one enterprise customer cutting review cycles by roughly 75 percent; a 2025 Points of Light case study describes a deployment at an unnamed 550,000 member Pennsylvania plan on a defined sample of outpatient cases. Scale of use, plans covering more than 50 million lives, is not evidence of benefit.

No method, denominator, reference standard or period accompanies any figure, and accuracy is not defined, whether as agreement with a nurse reviewer, with the final determination, or on appeal. The KLAS work is cited by the vendor and its method is not published.

General assurances, and one sentence that raises the training question rather than answering it. The Florence page says feedback from language model checks and human review flows back into the system so Florence keeps improving with every case, and that each plan's decision standards are encoded and version controlled. That establishes that production cases feed the system; it does not say whether that means configuration, policy and prompt updates or model training on member records, or whether anything learned from one plan reaches another. The privacy policy's de identification and research language covers website and application data, and it expressly excludes protected health information processed for plans.

Missing are a retention period after a determination, a deletion commitment, a statement on training use and a de identification step for member records. The immutable ledger records every access from record to outcome, which is an access control, not a retention answer. No safety engineering is published beyond pre deployment evaluation on synthetic data and escalation when criteria are not clearly met.

Status is stated, and nothing readable supports it. The footer carries a HIPAA link to the company's trust center, the Florence page says the system is HIPAA compliant, and the privacy policy, effective 13 August 2026, says that when providing the application to health plan customers the company processes protected health information on their behalf, restricted by its agreements with them, which describes a business associate role in the company's own words. The same sentence takes that processing out of the policy's scope, so the only published privacy document covers website visitors and application users, not member records. It also names co-helm.com and platform.co-helm.com as the site and application it covers.

No business associate agreement, subcontractor flow down or breach timetable is published outside the trust center, which may hold them. A plan is a covered entity, so live customers very likely have executed agreements; members whose records are reviewed are not party to them.

SOC 2 Type II and HITRUST e1 are named as the company's own, with the artifacts in a trust center. The site footer carries Trust Center, SOC II, HIPAA and HITRUST e1 links to trust.anterior.com, and the Florence page lists SOC 2 Type II and HITRUST e1 under security and says the system is HIPAA, SOC 2 and HITRUST compliant. Whether the portal releases reports openly or on request is not stated on the public pages.

The SOC 2 period and auditor are not named, and HITRUST e1 is the essentials level of that framework, a one year assessment of a foundational control set rather than the risk based r2. The same page also says the system is built to SOC 2 requirements, a weaker phrase than certified, beside the certification claims.

No device claim is made, and the product is scoped to administrative and population health functions. Utilization management recommendations go to plan clinicians, and the July 2026 Stellarus announcement says licensed clinicians make every final determination and the AI will never deny care; payment integrity reviews claims against the chart and contract; HEDIS work closes care gaps; and risk adjustment surfaces suspected HCCs and over or under coded diagnoses with each cited back to the source chart for review, a documentation function with a reviewable basis rather than a claim to find disease to treat.

The law that governs is coverage law: federal prior authorization and interoperability requirements, state utilization review statutes, state laws restricting automated coverage denials, and accreditation. The Florence page says the system is built to NCQA and URAC requirements and is compliant by construction; built to is not accredited, and no URAC or NCQA accreditation held by the company is named on the home, Florence or privacy pages. Where a plan delegates review, its obligations stay with it, and accreditation the vendor does not hold leaves the delegation oversight with the plan.

An evaluation practice is described, and no framework, result or subgroup analysis is published. The Florence page says synthetic data and large scale evaluation catch issues before deployment, that feedback from language model checks and human review flows back in production, and that each plan's policy interpretations are held as version controlled standards. That is a practice stated as a feature, with no named body, metric or result behind it, and nothing on governance appears on the home or privacy pages.

The stakes are specific to utilization management. Performance can vary by service line, by documentation quality and by the population a provider serves, and documentation completeness tracks practice resourcing, so a system that routes more incomplete records to review can slow access for patients of under resourced practices even when it never denies. No result by service line, request type, provider type or population is published, and no appeal overturn rate for cases the system touched.

Mechanisms exist to challenge an output, and nothing stands behind it. Florence attaches its reasoning and supporting evidence to each outcome, naming the record findings and the specific policy criteria applied, and writes every action from record access to final outcome to an immutable ledger the company describes as auditable for appeals and regulatory review. Cases that do not clearly meet criteria go to a qualified human with a sourced evidence packet, and the July 2026 partnership announcement with Stellarus states that the AI may recommend approval or route a case for review but will never deny care.

No warranty, indemnity, accuracy commitment or remediation obligation is published, and no terms of service are linked from the footer. The 99.24 percent clinical accuracy the company attributes to validation by KLAS Research is published without a method or denominator, so it is not an error rate a buyer can test. A provider or member who disputes an outcome goes through the plan's appeal process, which the product supports but which runs through the customer, not the vendor. No fee at risk against a measured result is published.

Named plan systems and a published list of standards. The Florence page says the system uses FHIR internally and ingests X12 EDI, HL7 v2, C-CDA, NCPDP, SFTP extracts, proprietary feeds and unstructured PDFs, faxes and scans, with documented SDKs and APIs for common plan systems and bespoke paths where needed; accepting fax and scan is what reaches a whole provider network. The company's announcements name an integration with HealthEdge's GuidingCare care and utilization management platform and a July 2026 partnership integrating Florence with Stellarus's AtlasIQ platform for real time prior authorization.

The direction and depth of each integration are not described, no marketplace listing or integration documentation is published, and the standards based prior authorization interface plans must now run is not named as supported, so whether the product helps a plan meet that requirement or sits beside it is not answered.

A single hosted service with its location implied. Florence runs as background agents or under human in the loop review, its interface components run standalone or embedded in a plan's existing systems, and the application sits on the company's own platform domain; the company is based in New York and sells to United States plans. The only transfer statement published, in the privacy policy, says service providers may operate in the United States and other countries, and it is scoped to website and application data, not member records.

No cloud provider, region, tenancy model or residency commitment for protected health information is published, though the trust center may hold more. The company embeds its own clinicians with each plan to configure the system, a delivery approach rather than a deployment option, which makes vendor access to live records during tuning a direct question.

No price is published, and how the product is sold is discoverable. Anterior sells four named programs to health plans (Triage for prioritizing AI initiatives, Flow for clinician productivity, Relay for end to end automation such as prior authorization, and Steward for work across departments), delivered with clinicians the company embeds in each plan, and it also reaches plans through partners: the Stellarus AtlasIQ platform and utilization management organizations such as MedWatch and HealthHelp. Value is framed as lower medical and operating expense, and the route in is a conversation with the sales team.

What a buyer cannot establish before contact is what drives the cost: per covered life, per case reviewed or per determination. The difference matters in utilization management, where a fee per determination rewards volume.

Named across payer workflows, with validation behind part of it. The home page lists utilization management, HEDIS gap closure, payment integrity, risk adjustment, policy management, and appeals and grievances, and the Florence page adds case management; the programs are sold to health plans and to organizations that run clinical review for them. MedWatch, a national utilization management organization, is the named deployment with figures across utilization management, care management and payment integrity; the company's announcements name Geisinger Health Plan as a customer, HealthHelp as a utilization management partner, and Stellarus, spun out of Blue Shield of California, as a platform partner for real time prior authorization. The HEDIS and risk adjustment functions carry the value based care work: gap closure to lift Star Ratings, and suspected HCCs surfaced from charts with each cited back to the source.

No result is published for HEDIS, risk adjustment, policy management or appeals, the service lines within prior authorization are not enumerated, and coverage is payer side by design.