Authenticx

The models are the engine of the core platform, and a services line sits beside them. Authenticx says it runs more than 20 proprietary healthcare models across its Business Insights, Quality and Coaching, and Safety and Compliance products, covering conversational friction (its Eddies), agent quality, sentiment, topic discovery, compliance and adverse event signal detection, and without them analyzing every interaction is impossible. But the company also sells Client Success, Insights Services workshops led by its experts, and a Conversation Analysis service, customers describe round the clock access to a personal consultant, and the Accuracy and Reliability page says human expertise, the customer's and the company's, ensures insights are interpreted correctly.

Without the models, the analysts and consultants remain with far less to work from, and part of what is sold is their interpretation. The record's earlier statement that no services layer exists no longer matches the vendor's own navigation.

Nothing acts on a patient, a review structure is described, and one part is missing. The platform reports to leaders, guides agents during calls and scores interactions afterward; the Accuracy and Reliability page says on shore teams review the AI continuously, that transcripts and summaries can be validated, and that because the AI is probabilistic, human expertise interprets its findings before anyone acts. The same page claims 100 percent visibility with zero guesswork, analyzing every interaction rather than a sample. That is a claim about coverage, not correctness, and it sets the scope the oversight has to cover: every adverse event in every call.

The missing part is what happens after a flag. The home page speaks of streamlining the escalation of patient safety outcomes, and no threshold, timeline or owner of the decision that a flag is reportable is published on the home, Accuracy and Reliability or Privacy and Security pages. The second question is employee scoring: agent quality is a model output, and nothing says whether an agent can see or contest a score or whether it feeds employment decisions.

The models are named by function and nothing about them is identified. The company says it runs more than 20 proprietary healthcare models, covering conversational friction, agent quality, sentiment, topic discovery, compliance and adverse event signals, trained only on real healthcare conversations labeled by experts, with transcription and a generative summarization component producing the transcripts and summaries customers review. It cites 97 percent AI accuracy.

What is missing is the architecture, whether any component is built on a licensed foundation model, how the training data was characterized and sourced, and a version and update practice that tells a customer when a model changed. Proprietary is asserted rather than explained, and the accuracy figure names no task or method. Naming which component generates text still helps a reviewer, because summaries can state things a caller never said while classifications can only mislabel.

The host is named, and the model layer is asserted as proprietary without saying what that covers. The Privacy and Security page says the platform runs on Microsoft Azure across multiple data centers, with each client in a dedicated resource group and identities managed in Azure's directory service, and the Accuracy and Reliability page says the company runs more than 20 proprietary healthcare models. The company does not say whether the transcription and generative summarization components are built in house or licensed, or from whom. The terms let the company engage subcontractors at its discretion without naming any.

The named host matters for residency and breach scope; it does not say which model sees a call. If every model is built and run in house on Azure, the chain is short and a list would show it.

Vendor reported figures and a scale claim, with no method a reader can test. The Accuracy and Reliability page cites 97 percent AI accuracy, more than 20 proprietary models and more than 500 million healthcare conversations analyzed, and customer stories report 58 percent fewer calls from health care professionals and 70 percent faster copay enrollment at pharmaceutical customers. A January 2024 payer story reports a top five Medicare plan finding friction in 22 percent of member calls and, after fixing its portal, 26 percent more portal traffic, over 300 agent hours and nearly $10,000 saved a month, and 6 percent less caller friction in three months, with no CAHPS or Star Ratings result. The same story reports a Fortune 500 payer's AI quality evaluations rising from 69 to 99 percent agreement with its human analysts, across 4.9 million conversations in 2023. The company is the claimed source and the customers are unnamed, and scale of use is not evidence of benefit.

Missing are a task or reference standard for the 97 percent, a sample size behind the 99 percent agreement, baselines or periods for the customer results, and any independent or peer reviewed evaluation. Adverse event detection sensitivity, where a miss carries a reporting consequence, is not published.

A categorical commitment is published, without the retention schedule behind it, and it sits beside a training claim the company does not reconcile with it. The Privacy and Security page says data is processed solely to deliver insights for the customer, that sensitive healthcare data is never sold or used to train the company's models, and that only the data needed for the service is collected and retained. The terms add export of customer data on request within 30 days of termination, and let the company use usage data (query logs and metadata) to improve its products.

No retention period is published for recordings, transcripts or derived results, and no deletion commitment follows export. The Accuracy and Reliability page also says the models are trained on real healthcare conversations labeled by experts, and does not say where those conversations came from or whether de identified customer calls fall outside what the company treats as sensitive. No safety engineering is published beyond human review.

Business associate status is stated and supported, and the agreement is not published. The Privacy and Security page lists the company as HIPAA compliant and describes purpose limitation, data minimization, a commitment not to sell or train on sensitive healthcare data, per client isolation and audit logging. The terms (last updated 15 February 2024) require a business associate agreement or a data use agreement before a customer uploads HIPAA data, incorporate that agreement by reference once executed, and refer to a data privacy addendum and a security addendum.

What keeps it out of A: none of the agreement, the data use agreement or the two addenda is published, so whether the agreement binds subcontractors and what its breach timetable is cannot be read before contracting. The terms also state the company has no liability for HIPAA data uploaded without an agreement in place, which puts the burden of executing one on the customer.

A recognized certification is named as the company's own, without the artifact. The Privacy and Security page lists SOC 2 Type I and Type II under the AICPA mark, alongside HIPAA, GDPR and CCPA, and describes controls in detail: least privilege access in dedicated client resource groups, multi factor authentication for internal services, single sign on, audit logging of access and data activity, background checks and security training, encrypted daily backups and annual disaster recovery testing. The company announced a SOC 2 Type 1 examination in February 2021; the page now names Type II.

The Type II period, auditor and trust service criteria are not named on the Privacy and Security page, no report or trust portal is offered to open, and no penetration testing cadence is stated.

No device claim is made, and the regime that applies is pharmacovigilance and complaint handling rather than device clearance. The platform analyses conversations for patterns, agent quality and compliance, and its Safety and Compliance product detects adverse event signals in the calls a manufacturer's contact center handles; the recommendation it produces is to review and escalate a possible safety event, not to change a patient's treatment. Drug manufacturers must collect and report adverse events they become aware of, and device makers must handle complaints and report device problems, so software that surfaces those signals sits inside a regulated reporting chain.

The questions that follow are detection sensitivity, timeliness, documentation and who is accountable for a missed signal, and none is answered on the home, Accuracy and Reliability or Privacy and Security pages. The home page calls the safety models highly accurate and industry leading without a figure.

A governance practice is described, and no evaluation or subgroup result is published. The Accuracy and Reliability page says the models are trained only on healthcare conversations labeled by experts, that the AI is built by on shore teams with continuous human review, that outputs are monitored for consistency across channels and teams, and that the AI is probabilistic so human expertise interprets what it surfaces; it cites 97 percent AI accuracy with no method, task or denominator. That is a practice stated as a feature, with no named body, framework, metric definition or result behind it.

Three exposures follow from what the product does. Speech and language performance varies with accent, dialect, first language, age and emotional state, and nothing is published by group. The platform scores agents against rubrics, so any variation in how well the models understand a caller carries into the evaluation of the employee who took the call. And adverse event detection fails silently: a missed signal is invisible to everyone, including the manufacturer whose reporting duty does not shrink because software missed it.

Mechanisms exist to check an output against its source, and the terms stand behind nothing. The Accuracy and Reliability page says every interaction is analyzed with transcripts and summaries a customer can validate and share, that leaders can zoom from enterprise patterns into individual conversations, and that system activity is logged; that is traceability from a finding back to the call. The terms (last updated 15 February 2024) provide the services as is and state that the company makes no warranty that they will be accurate, complete or error free, cap the company's liability at the fees paid in the prior 12 months, and have the customer indemnify the company for claims arising from customer data. The only remedy the company takes on is for intellectual property claims.

No error rate with a method and denominator is published: the 97 percent accuracy figure names no task, and a 2024 customer story's 99 percent agreement between AI and human quality evaluations gives no sample size. No route exists for a caller or agent who disputes what a model concluded about them, and for adverse event detection nothing says what the company owes when a signal is missed.

Named contact center and CRM systems, with the depth of each left open. The Integrations page says the platform connects through APIs to a customer's CRM, contact center and integration stack, and shows Five9, Genesys, Amazon Connect, NiCE, Vonage, Salesforce, AWS, MuleSoft, Twilio, Qualtrics and ServiceNow as available integrations, with a technical guide offered; the agent assist product is distributed through the Salesforce marketplace. The target is deliberately the contact center and customer relationship stack rather than the clinical record, which is right for this product: the conversations it analyzes happen in telephony and messaging systems, not in a chart.

The page names the systems without saying which are read only and which receive results back, and none of the published material ties a production deployment to a named integration.

A single hosted service on a named cloud, with isolation described and residency not committed. The Privacy and Security page says the platform runs on Microsoft Azure across multiple data centers with failover, that each client operates in a dedicated resource group with least privilege access, that backups are encrypted daily, and that disaster recovery and restoration are tested annually. The same page lists GDPR alongside HIPAA and CCPA. No region or residency commitment is stated on the home, Accuracy and Reliability or Privacy and Security pages, and the terms leave data localization obligations with the customer.

Where recordings and transcripts rest is implied by a United States company on Azure rather than committed, and whether any European data is processed elsewhere is not said. For a product whose value comes from a long history of conversations, how long that store persists matters as much as where it sits.

No price is published, and how the product is sold is discoverable. Authenticx sells a platform of three products (Business Insights, Quality and Coaching, and Safety and Compliance) with integrations, alongside services (Client Success, Insights Services workshops and Conversation Analysis), to pharmaceutical and life sciences companies, device makers, health insurers and providers. The terms describe subscriptions under order forms with software services, managed services and technical services, and the route in is a demo.

What a buyer cannot establish before contact is the unit of charge: per seat, per conversation analyzed or per program. Its last announced round was a $20 million Series B in January 2023, and because the product's value grows with the conversation history it holds, switching later means losing that baseline.

Coverage is named by industry, with validation behind part of it. The company serves pharmaceutical and life sciences companies (patient access, adherence and safety oversight), device makers (onboarding, troubleshooting and support), health insurers (Star Ratings, member experience and audit readiness) and providers (access, care coordination and support services), across telephony, chat, messaging, email and CRM notes. The insurer work ties the product to value based care: a 2024 customer story describes a top five Medicare plan tuning the company's models to member conversation topics that track negative CAHPS responses, to raise its CAHPS scores and Star Ratings, which drive Medicare Advantage quality bonus payments. The pharmaceutical segment carries most of the evidence (58 percent fewer calls from health care professionals; 70 percent faster copay enrollment), and insurer stories report portal traffic, agent time and call savings.

Customers are unnamed or identified by first name and sector, the Medicare story reports operational results but no CAHPS or Star Ratings change, the provider segment is described rather than evidenced, and geography is the United States.