Luma Health

A decade old patient access platform that has put language models underneath itself. Scheduling, reminders, waitlists, intake, payments and record system connectivity came first; Spark, the generative core, was announced in October 2024, and Navigator, the conversational agent, and Fax Transform, now the Document Flow Agent, were its first products. The company says it used machine learning and language processing for years before that. The Summer 2026 release added kiosk and QR code check in, an automated arrival queue and card reader support, which are platform features rather than model work.

The model layer does real work inside the record system: the conversational agent verifies callers and cancels or books appointments directly in the record, the document agent turns inbound faxes into structured referrals and outreach, and a 2026 outbound agent contacts patients to close care gaps. Take the models away and a working patient engagement platform remains.

High autonomy with an explicit statement of who sets the bounds. The conversational agent verifies a caller's identity, retrieves appointments and cancels or books directly in the record system without a person in the loop; the company's engineering lead has described it as several cooperating agents, each a model with a prompt and a set of tools. Customers assemble agents from individual skills through a configuration interface, and the agent chooses each patient's next step from options the organization has predefined, so the decision space is bounded by the customer. The company also publishes material intended for a customer's AI governance committee.

Two things are missing. Nothing published describes what happens when a caller describes an emergency or acute distress, which every product answering a health system line has to answer. And the outbound agent released in 2026 decides who needs contacting to close care gaps, with no published criteria or review step. No figure shows what share of calls reach a person.

The model suppliers are named and the version discipline is not. The company's October 2024 Spark announcement describes multi model, zero retention generative AI including OpenAI's GPT-4, Anthropic's Claude Opus, Deepgram and others, fine tuned for healthcare, used to understand natural language on phone and text, parse faxes and forms, and provide next steps. The company's engineering lead has described the conversational agent as several cooperating agents, each combining a model, a prompt and a set of tools, with a worked example of a cancellation executed in the record.

Provenance is not performance. No accuracy, containment or task completion figure, evaluation method or model documentation is published, the named models date from 2024 and "and others" is open ended, so which versions run today is unstated, and nothing describes how model changes are validated or announced.

The model providers and the host are named; a full list and the terms that bind them are not. The company's October 2024 Spark announcement names OpenAI's GPT-4, Anthropic's Claude Opus and Deepgram among the models it uses, described as zero retention and fine tuned for healthcare, and the security and trust page names AWS as the infrastructure. That tells a buyer which outside parties receive patient text and audio.

The gaps are the rest of the chain and a statement that sits awkwardly with it. "And others" leaves the list open, no subprocessor list is published, and the security page says customer and patient data is not made available to any of the company's partners or service providers, which cannot be literally true of the named model providers; the page does not explain the distinction it has in mind.

Named and specific figures, all vendor produced. The home page names the University of Arkansas for Medical Sciences, whose chief clinical access officer is quoted: its access center handles 1.2 million calls a year, the conversational agent automates 95 percent of after hours calls, and after hours cancellations alone save more than 800 hours a year on a volume of 10,000 calls, running against Epic. The home page also reports averages of 61 days less time to care, two to three hours of manual calls saved daily and a 47 percent increase in revenue, and the company has reported AI workflows at more than 50 health systems saving more than 2 million staff hours in 2025.

The UAMS figures come with a volume behind them; the averages and the 2 million hours carry no customer count, baseline or method. No peer reviewed or independent evaluation appears, and nothing measures how often the agent completes a request correctly rather than how often it completes one.

Categorical commitments on training and sharing, with one statement that needs reconciling. The security and trust page says the AI enabled products operate under a zero retention model and do not store data long term, that customer and patient data is not made available to other customers or to any of the company's partners or service providers, and that customer data is not used to train, retrain or improve any model.

Two gaps remain. The company's 2024 Spark announcement names OpenAI's GPT-4, Anthropic's Claude Opus and Deepgram among the models used, so patient text and audio evidently reach those providers during processing; the statement that data is not made available to any service provider needs to be read as meaning something narrower, and the page does not say what. And not storing data long term is not a retention period: nothing states whether calls are recorded, how long audio and transcripts are kept, or what the platform itself retains outside the model boundary. Nor is it stated whether callers are told they are speaking to software.

HIPAA compliance is stated and supported by a substantive service level description, with the agreement unpublished. The security and trust page says all of Luma's software and company processes are HIPAA compliant, that the company uses customer data only to provide contracted services and never shares, sells or uses it for advertising, that AI products operate under a zero retention model, and that customer data is not used to train any model; it also lists HITRUST r2 certification. For its advertising product the company has said protected information is never provided to the search provider and stays within its environment and the record system.

No business associate agreement, tier or subprocessor list is published, so a counterparty can read what the company does but not the document that binds it. Whether the agreement reaches the model and speech providers, or data handled by staff outside the United States, is not stated.

Five credentials named precisely, and no report a buyer can open. The security and trust page lists HITRUST CSF r2 certification, TX-RAMP Level 2, ISO 27001:2022, participation in the EU-U.S. Data Privacy Framework, and an annual SOC 2 Type II attestation, with accredited third party audits at least yearly, annual external penetration testing by a third party, monthly vulnerability scans, access reviews every 60 days and a documented secure development process.

The weak point is retrievability. There is no trust center and no report or certificate to open; the Data Privacy Framework listing can be checked on the Commerce Department's site, and the rest must be requested. Whether the conversational and document agents sit inside the audited scope is not stated.

No device clearance, authorization or submission is claimed and none would be expected. Scheduling, reminders, intake, document processing, payments and check in are administrative, and the outbound agent released in 2026 contacts patients to close care gaps, which is scheduling and outreach rather than a clinical recommendation.

The regimes that do apply sit elsewhere. Automated outbound calls and texts carry telephone consumer protection obligations, and the outbound agent increases that exposure. Payment collection and card readers bring payment rules. The advertising product launched in August 2026, which links search advertising spend to booked appointments for organizations on Epic, sits in an active enforcement area for health information in online advertising; the company's statement that protected information never reaches the search provider is the right thing to have published, though which identifiers cross the boundary is not specified.

How outbound consent is captured is not described, and neither is the agent's handling of a caller describing an emergency.

The company says its AI enabled products are built to be ISO 42001 compliant, which is not the same as certified. The claim appears on the security and trust page, ISO 42001 being the AI management system standard, and the company publishes material designed for a customer's own AI governance committee. Its other credentials are stated precisely, with a standard version and an attestation type. No completed ISO 42001 audit is mentioned.

The bias gap is the larger issue. The conversational agent answers calls in multiple languages across hundreds of organizations, and no performance is published by language, accent or caller group.

The platform shows what the agents did, and nothing published stands behind it. The conversational agent returns an updated schedule to staff after acting in the record, the collaboration hub gives staff visibility into agent conversations with translation, the agent's choices are bounded by options the organization predefines, and a public status page reports uptime and incidents against a 99.9 percent target. A customer can therefore see and correct what an agent did.

Nothing on the site puts an accuracy or task completion figure, warranty, indemnity or remediation commitment behind the agents, though the terms may. The agents write directly into the record, so a cancellation made for the wrong patient or a booking in the wrong slot is found only when staff look. Beyond the updated schedule, nothing describes how staff are alerted to an action that may need review.

Bidirectional integration with named record systems, documented and running in production. The site states bidirectional integration with all major record systems and publishes integration pages for Epic, Oracle Health, MEDITECH, eClinicalWorks, athenahealth, NextGen, Greenway and Nextech; the home page links to Greenway's own patient engagement page, where Luma is offered within Greenway's product. The company's 2024 Spark announcement names the same bidirectional integrations, and the company has described building its record connectivity itself rather than through a third party integration platform.

Production evidence is named: the University of Arkansas for Medical Sciences runs the conversational agent against Epic across an access center handling 1.2 million calls a year, with cancellations written directly into the record. The agents write as well as read, and the document agent turns inbound faxes into structured downstream work with the loop closed back to the referring provider. Which integrations write back for each module, and how record system upgrades are handled, are not described.

A single cloud service on a named provider, with the region implied rather than committed. The security and trust page says infrastructure runs across multiple AWS availability zones with real time replication to a 99.9 percent uptime target, is managed as code, can be deployed into any AWS data center if needed, and is tested for continuity annually; a public status page shows uptime and incidents. The company holds TX-RAMP Level 2, a Texas state cloud authorization, and participates in the EU-U.S. Data Privacy Framework.

No hosting region or residency commitment is stated, and the ability to deploy into any AWS data center leaves location open. Retention for platform data is not stated beyond the zero retention model for AI traffic. The company has research and development centers in Brazil and Europe, and nothing says whether staff there can reach patient data or under what controls.

The shape of the offer is enumerated and nothing is priced. The site lists the Patient Success Platform module by module (referrals, scheduling, waitlist, reminders and recalls, broadcast messaging, intake and forms, registration and preparation, eligibility and verification, payments, feedback, analytics, the collaboration hub, the conversational agent, the document flow agent, patient pipeline and novel workflows) and segments buyers into enterprise health systems, specialty groups, regional and rural care and primary care, with an implementation and support page and a demo route.

What no page states is the unit of charge, whether AI products are priced separately from the platform, implementation fees or the contract term. The cost drivers a buyer can anticipate are record system integration scope, workflow design and change management for scheduling rules. The last disclosed funding round was a $130 million Series C in November 2021.

Broad across organization type and channel within one country. The company describes serving more than 600 health systems, integrated delivery networks, federally qualified health centers, specialty networks and clinics, and segments its buyers into enterprise health systems, specialty groups, regional and rural care and primary care. The channel span covers voice, text, web self scheduling, inbound fax, kiosks and payments in one platform, which reaches older and rural patients and referring practices that still send paper, and the conversational agent operates in multiple languages with translation for staff.

Validation is named for an academic medical center, UAMS, on the home page. That covers one setting in depth, and the deployed footprint is entirely in the United States although the company has research and development centers in Brazil and Europe. No rural, primary care or specialty deployment is named.