AKASA

Generative models are the engine of each AKASA solution, and the platform around them carries value of its own. The company says its healthcare-specific language models are trained on clinical and financial data, 43 million clinical documents and counting, and fine-tuned for each health system on its own data. They power coding and documentation suggestions, a research assistant for revenue cycle staff, and the automated checks behind Auth Status and Claim Status.

The platform page lists four other components beside the models: an in-house team of revenue cycle experts averaging more than 15 years' experience who can take over whole workflows, standards-based EHR integration, production reporting with peer benchmarking, and security. Take the models away and a managed revenue cycle service with integration and reporting remains, so the models are the engine of the core modules rather than the whole of what is sold.

Review by people is described on each product, and the point at which work proceeds without one is not. Auth Status is sold as an AI plus expert-in-the-loop service: AKASA checks the status of pending authorization requests on payer portals and documents the results, and the Montage Health case study says accounts needing intervention are routed to staff or AKASA work queues.

Authorization Advisor, as the company describes it, pulls suggested documents and justifications into a sidebar for the prior authorization team to evaluate before submission. The platform page offers two modes, AKASA's experts working with the automation to remove whole workflows or the technology placed in the customer's own staff's hands, and the 2021 security overview says tasks the automation does not recognize go to internal experts.

There is no published confidence threshold deciding which work goes straight through, no figure for how often experts intervene, and no account of what happens after a wrong status or code reaches the record. The product pages offer accuracy, speed and completeness as benefits; none says the output is error free.

The approach is named, and neither the base model nor the version discipline is. AKASA says it has built healthcare-specific language models trained on clinical and financial data, 43 million clinical documents and counting, and fine-tunes a tailored model for each health system on that organization's own data.

The home page says the company improved the generative AI that powers ChatGPT by optimizing it for healthcare, which describes the technology family without naming the model underneath. The March 2024 Authorization Advisor release says the tailored models are 38 percent more accurate than GPT-4 at selecting prior authorization documents, with no test set, measure or method. Status automation is described in the case studies as AI and machine learning decisioning with experts in the loop. No architecture, model version, update cadence or notice of model change is published.

The host is named and the model layer is not. AKASA's August 2022 HITRUST announcement says the platform is hosted on Amazon Web Services, and the 2021 information security overview describes AWS identity and access controls. No model provider or base model is named: the home page says the company improved the generative AI that powers ChatGPT by optimizing it for healthcare, and the March 2024 Authorization Advisor release compares its models with GPT-4, but neither says whether an OpenAI or other external model sits underneath, or on what terms.

No subprocessor list is published. The data at issue is complete clinical and financial documentation from whole health systems, and the destruction commitment in the 2021 overview is the only published term governing it once it arrives.

Named deployments with dated figures and some method, strongest on the status products. AKASA's 2023 Montage Health case studies give volumes and periods: claim status automation built in four months, about 3,000 claims across five payers in the first three months live, more than 23,000 in six months and about 80,000 in a year, with 300 or more staff hours saved a month and a 13 percent fall in accounts receivable days that the study says the work contributed to; authorization status checks across nine payers, 5,663 statuses in a year and a 22 percent smaller Epic authorization work queue.

For Methodist Health System the figures are 71.3 percent of accounts cleared from staff queues and an 86 percent cut in time on claim status checks. No baselines, no comparison period and no independent measurement are given. The 2024 Authorization Advisor launch claimed up to 50 percent less time on prior authorization and models 38 percent more accurate than GPT-4 at selecting documents, with no test set or method. Cleveland Clinic is deploying the coding and documentation tools, with no results published. The stated client base of 500 hospitals is scale of use, not evidence of benefit.

A retention commitment is published, and the training question is only half answered. AKASA's 2021 information security and privacy overview says it destroys all sensitive data after a customer contract ends, or during an engagement on the customer's preferred retention schedule.

Training use is described as design rather than limited: the company trains healthcare-specific models on clinical and financial data, 43 million clinical documents and counting, fine-tunes a model for each health system on its own data, and the overview says tasks resolved by its experts feed further model training. Not stated: whether a customer's documents join the shared corpus behind other customers' models, and anything on de-identification.

The data is complete clinical and financial documentation for whole health systems. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not described.

HIPAA status is stated and a service document stands behind it, while the agreement itself is not published. The platform page describes AKASA's security infrastructure as HIPAA compliant. The website privacy policy, last updated in March 2023, says it does not apply to personal information processed for customers, which is governed by customer contracts and subject to the customers' own privacy policies and notices of privacy practices; that contract is where to ask.

AKASA's 2021 information security and privacy overview reaches the service: it describes how customer data is collected and transferred, logging of every access to protected health information, and destruction of sensitive data at contract end or on the customer's retention schedule. No business associate agreement, its terms or the subcontractors it binds are published.

AKASA presents its certifications as its own, and none is published as an artifact with a current date. The platform page says AKASA is SOC 2, NIST 800-53 and CIS certified, and a 2022 company blog specifies SOC 2 Type 2. An August 2022 press release announced HITRUST r2 certification for the platform, hosted on Amazon Web Services. An r2 certification runs for two years, the current platform page no longer names HITRUST, and no renewal is published.

NIST 800-53 and the CIS benchmarks are control frameworks rather than audits a vendor is certified against. The platform page's link to the company's data protection page returns an error, and no trust portal or report is available without asking.

These are administrative and financial products, and they carry no device claim. AKASA's tools suggest codes and documentation improvements, check authorization and claim status on payer portals, select documents for prior authorization requests and answer revenue cycle staff's questions; none diagnoses, treats or recommends care.

The rules that govern this work are real and fall on the health system, which signs and submits the claim: CMS coding and billing rules, the False Claims Act and OIG compliance guidance, and each payer's requirements for what a prior authorization request must contain. Documentation queries that prompt a clinician to add specificity also sit under industry guidance against leading queries. AKASA publishes a comparative accuracy figure for document selection but no coding accuracy against a reference standard.

Performance is monitored and reported, and nothing is published on how model behavior is evaluated across the work it does. The platform page offers production reports and benchmarking against peer organizations using AKASA, and the 2021 information security overview describes persistent performance monitoring, with tasks the automation does not recognize sent to internal experts.

The coding product is sold to surface quality and compliance risks as well as missed coding opportunities, and Authorization Advisor's suggestions carry justifications; both let a person check an output, and neither is an evaluation. No governance framework with a named process, no results by payer, service line, document type or encounter type, and no bias or error analysis is published. For coding and documentation suggestions the groups most at risk are the encounter types and specialties a tuned model sees least; for status automation, the payers whose portals change most.

Mechanisms exist to catch a wrong output, and nothing stands behind one that gets through. The Montage Health case study says status results are returned into the record with detail and accounts needing intervention are routed to staff or AKASA work queues; Authorization Advisor's suggestions carry justifications for the prior authorization team to evaluate; and the 2021 information security overview says every access to protected health information is logged and original content is kept when a data object changes.

No warranty, indemnity, service credit tied to output quality or remediation commitment is published, and no error rate with method and denominator. The claim that AKASA's models are 38 percent more accurate than GPT-4 at selecting prior authorization documents is a comparison with no test set, not an error rate. On coding the exposure sits with the health system, which signs and submits the claim.

Epic is named in production, with status results written back into it. AKASA's Montage Health case studies describe claim and authorization status automation working alongside Epic, returning detailed status information into the record to drive follow-up, and reducing an Epic authorization work queue by 22 percent.

The platform page says integration with record systems is standards based, through an API, EDI or other means, and the company's own marketplace copy for Authorization Advisor says its sidebar reads from the record over FHIR and works with most record systems, naming Epic and Cerner. No marketplace listing or integration documentation is published, and nothing describes what is written into the record beyond status results, or how depth differs on Cerner, MEDITECH or other systems.

A single hosted platform on a named cloud, with the location implied rather than committed. AKASA's August 2022 HITRUST announcement says the platform is hosted on Amazon Web Services, and the 2021 information security overview describes data moving by SFTP or VPN and in some cases encrypted email, encrypted in transit and at rest. No AWS region, residency commitment or tenant isolation model is published.

The website privacy policy says information may be processed anywhere in the world, but it covers website visitors, not customer data. Part of the product touches the customer's own environment: the 2021 overview describes remotely installing proprietary software at the start of an engagement to analyze how staff work, and Authorization Advisor runs as a browser sidebar reading from the record.

No price is published, and how the product is sold can be established before contact. AKASA sells to hospitals and health systems by solution, each with its own demo request. The platform page describes a dedicated account manager, production reporting, and the option of AKASA's own revenue cycle experts taking over workflows rather than staff using the tools, a service component that will shape the cost.

No unit of charge, fee basis or implementation fee is published, and nothing says whether any fee depends on recovered revenue, yield or coding level. The Montage Health case study reports a four-month build before go-live.

Coverage is named at the level of the health system, with validation behind part of it. AKASA sells to hospitals and health systems, states a client base of 500 hospitals across all 50 states, and covers coding and documentation integrity before billing, prior authorization status, claim status and a research assistant for revenue cycle staff.

Validation sits behind the status products: Montage Health's hospital and medical group for claim and authorization status, with volumes and periods, and Methodist Health System in Nebraska for claim status. Cleveland Clinic is deploying the coding and documentation tools across its US locations with no results published, and nothing states whether coding covers inpatient, outpatient or professional encounters, or which specialties.

In prior authorization AKASA works on the requesting side: Auth Status checks pending requests on payer portals and logs the result, and Authorization Advisor, launched in March 2024, assembles documentation for submissions, though it no longer appears among the listed solutions.