Banjo Health

The CARE engine does the substantive work: extracting structured facts from faxed and digital requests, reading prescriber notes from the EHR, and converting narrative clinical guidelines into executable decision trees. That guideline to decision tree transformation is the technically interesting step, since coverage policy written as prose is the actual bottleneck in prior authorization automation.

Two stated design choices work in the buyer's favor. Every AI generated recommendation carries detailed evidence and a clear explanation rather than a bare determination, and the company states staff intervention is required for exceptions once criteria are matched, placing humans on the hard cases. Held back from A because, unlike Cohere Health, the company does not publicly state whether automated determinations are limited to approvals.

In a pharmacy benefit context that distinction is decisive: an automated denial of a prescribed drug is a materially different act from an automated approval, and a buyer should establish in contracting which the platform is configured to issue.

The product's central output is an inspectable structure rather than an opaque score, which is materially more transparent than most of this lane. Composer converts narrative clinical criteria into decision trees that execute inside the workflow, so a plan can read the logic that will be applied to its members rather than trusting a model's judgement.

Held at B rather than A because the model performing the conversion is undisclosed, no accuracy figure is published for it, and no evaluation methodology was located.

The question this raises is the interesting one and it belongs in diligence. Not all coverage policy reduces cleanly to a tree. Guidelines routinely contain judgement language, clinically appropriate, medically necessary considering the totality of the record, and converting that into an executable branch requires an interpretive choice about what satisfies it. Establish who reviews the generated trees before they go live, whether the plan's own clinical staff sign them off, and what happens to a criterion the system cannot represent without narrowing it. That review step is where a coverage policy could quietly become stricter than the policy it was derived from.

A full tier certification provides externally validated controls over protected information, which is a stronger foundation than most records in this index carry and is the main reason this sits above the middle. Held below the top grade on a question the company's own description raises.

It states that its models continuously learn and adapt, which implies customer data improves the system, and nothing published says what data, whether it stays within a single plan's boundary, or whether learning from one payer's determinations informs models serving another.

For a vendor whose customers are competing plans and pharmacy benefit managers, that boundary is commercially as well as legally material: determination patterns are precisely what a competitor would want to know, and a model that improves by seeing how one plan adjudicates carries that improvement to the next customer whether or not anyone intended it.

A peer in this same segment answers the equivalent question explicitly with a per plan learning commitment, which establishes that the disclosure is available and unremarkable to make. The data itself is unusually rich, since the platform reads prescriber notes directly from the record system rather than working from a faxed form, so it holds clinical narrative written about a patient in support of a medication request, and no retention period was located for that content once a determination is made. Ask for the learning boundary in writing, and retention on prescriber narrative.

One substantial named implementation carries most of the weight: Navitus, described as the largest transparent pass through pharmacy benefit manager, owned by SSM Health and Costco and serving more than 18 million lives, expanded the deployment across employer and health plan lines including Medicare and Medicaid in December 2025. An expansion by an existing customer of that scale is stronger evidence than a new logo. Held back from A because turnaround and accuracy improvements appear as customer testimonial rather than measured figures with stated baselines.

HITRUST r2 certification provides externally validated controls over protected health information, which is a stronger foundation than most records in this index carry and is the main reason this sits at B.

Held at B rather than A on a specific unanswered question. The company states its models continuously learn and adapt. That implies customer data improves the system, and nothing published says what data, whether it stays within a single plan's boundary, or whether learning from one payer's determinations informs models serving another. For a vendor whose customers are competing plans and pharmacy benefit managers, that boundary is commercially as well as legally material.

The data itself is unusually rich. The platform reads prescriber notes directly from the electronic health record rather than working from a faxed form, so it holds clinical narrative written about a patient in support of a medication request. No retention period was located for that content once a determination is made.

No business associate agreement terms were located, but the HITRUST r2 certification incorporates the privacy and security rule requirements among the sources it harmonises, so the controls have been examined against them by an external assessor rather than asserted. The company also states adherence to health privacy and CMS regulatory standards in its partnership announcements.

That is the standard middle rung: demonstrable programme, undisclosed instrument.

One structural point worth establishing. The platform reads prescriber notes from the electronic health record, which means data originating at a provider organisation reaches a vendor working for the payer. The provider is a covered entity in its own right and is not the vendor's customer. Establish the authority under which that content moves, and whether it arrives through the plan's own agreements or through a separate route.

HITRUST r2 certification, the risk based two year assessment rather than the lighter e1, which is a substantive certification for a vendor handling prescriber notes and coverage decisions. Held back from A because no public trust center with control detail or audit dates was retrieved.

No FDA pathway applies and none is expected. Prior authorisation decision support for plans and pharmacy benefit managers sits outside Software as a Medical Device.

Graded B rather than C because the regime that governs is named and built into the product rather than left for a buyer to infer. The company positions explicitly around the CMS interoperability and prior authorisation final rule and its compliance deadline, and its partnership announcements state adherence to that framework and to health privacy standards.

The accreditation point is worth stating precisely so the grade is not misread. URAC and NCQA utilization management accreditation are the clearance equivalents in this lane, and neither was located here. But this vendor sells software to plans rather than performing delegated review with its own clinicians, which places it outside accreditation eligibility rather than short of it. That distinguishes it from Alaffia Health and Xsolis, both of which operate services arms and are therefore eligible.

The live exposure to watch is the growing set of state laws conditioning AI involvement in coverage determinations, on which no position was located.

No AI governance framework, bias evaluation or subgroup performance disclosure was located.

One published claim makes the gap concrete rather than generic. The company states its models continuously learn and adapt to evolving clinical and regulatory requirements. A prior authorisation system that learns from a plan's own history is learning from that plan's past determinations, which encodes whatever patterns already existed in them. Where those patterns were uneven, continuous learning preserves the unevenness and applies it faster and more consistently than the humans did.

That is the same concern this index records against Cohere Health, and it is unaddressed here too. Nothing published describes what the models learn from, whether historical determinations are used as training signal, whether approval and denial rates are monitored by drug class, plan line or member population, or whether anyone checks that automation has not shifted outcomes.

The pharmacy benefit context sharpens it. A denied medication authorisation is felt immediately by a patient standing at a counter, and the appeal burden falls on the prescriber.

The product's central output is an inspectable structure rather than an opaque score, which is materially more contestable than anything else in this lane. Narrative clinical criteria are converted into decision trees that execute inside the workflow, so a plan can read the logic that will be applied to its members, a reviewer can see which branch a case failed on, and a provider disputing a denial has a specific condition to argue about rather than a determination.

That is the right shape for a coverage decision, because the thing being applied is a policy rather than a prediction, and a policy should be readable. Held below the top grade because the model performing the conversion is undisclosed, no accuracy figure is published for it, no evaluation methodology was located, and no warranty, indemnity or remediation commitment attaches. The question that raises belongs in diligence and is the interesting one.

Not all coverage policy reduces cleanly to a tree: guidelines routinely contain judgement language such as clinically appropriate or medically necessary considering the totality of the record, and converting that into an executable branch requires an interpretive choice about what satisfies it. That review step is where a coverage policy could quietly become stricter than the policy it was derived from, without anyone deciding to tighten it. Establish who reviews generated trees before they go live, whether the plan's own clinical staff sign them off, and what happens to a criterion the system cannot represent without narrowing it.

Interoperability is central to the product rather than incidental. The platform reads prescriber notes directly from the electronic health record rather than relying on manual entry or fax and phone follow up, ingests both faxed and digital requests, and supports electronic prior authorisation, with one customer reporting market leading adoption of the electronic pathway after implementation.

A connectivity partnership extends reach further, linking into a network described as covering more than 13,000 connected payers and millions of participants including providers, electronic health record and practice management systems, and clearinghouses, with electronic attachment handling.

Held at B rather than A because no specific electronic health record platforms are named and no integration method is described for the note retrieval, which is the technically hardest and most consequential connection in the product.

No hosting provider, region, tenancy model or data residency commitment was located, and no subprocessor list is published. The HITRUST r2 certification is scoped to the platform and implies the infrastructure was assessed, but a certification scope is not a residency disclosure.

The tenancy question deserves a direct answer because of who the customers are. Plans and pharmacy benefit managers competing for the same employer and government contracts are running their clinical criteria and determination history on a shared platform. Establish what separates one customer's environment, criteria library and determination data from another's, and whether any model artefact spans them.

The generative component adds a second question. Composer uses generative AI to convert clinical policy into decision trees, so establish where that inference executes and whether a plan's proprietary coverage criteria are sent to an external model provider in the process.

No public pricing. Contact the vendor. Enterprise agreements with health plans and pharmacy benefit managers. Regulatory deadlines around CMS-0057-F are the stated adoption driver, which means buyers may be negotiating under time pressure; worth establishing whether pricing reflects compliance urgency or steady state operation.

Clearly bounded to payer side utilization management for health plans and pharmacy benefit managers, spanning prior authorization, appeals, grievances, and clinical criteria automation, with a pharmacy benefit emphasis that distinguishes it from the medical benefit focus of others in this lane.