ActiumHealth

Generative models conduct the communication and further models supervise it. Agents handle inbound and outbound patient contact across voice, text, short message and email, and a separate quality monitoring layer built on large language models reads the resulting calls. At the stated volume, 50 million calls and 100 million messages automated, no version of this product works without models doing the talking.

The earlier incarnation of the company was also model led, using record data to predict which patients most needed outreach, so both the product it was and the product it now is sit on models rather than on workflow.

High autonomy with the most interesting oversight answer in this segment, and it deserves to be named precisely. The agents talk to patients without a person present. What separates this record is that the company built a second artificial intelligence system whose job is to watch the first: a quality monitoring layer that analyses the calls its own agents handle each month, flags problematic interactions and allows intervention before a pattern spreads.

The reasoning its own operations lead gives is the notable part, that when new sites launch, features are added or the behaviour of an agent is changed, the organisation needs a full picture of patient experience rather than workflow measures alone. That is a description of behaviour drift and of the need to monitor for it, and this index has repeatedly recorded that no drift monitoring concept exists anywhere in this market. Held at B rather than A because the monitoring is retrospective, flagging interactions after they happened, and because no escalation criteria, emergency handling or disclosure practice is published.

Revised upward on 5 August 2026, and the reason is a document type this index almost never finds. The parent's trust centre publishes a named subprocessor list covering the cloud platforms and the model providers behind the product, spanning two cloud infrastructure vendors, two large language model vendors and a speech synthesis vendor.

Across more than fifty voice and conversation vendors reviewed for this index, nothing else has published its model supply chain at this level of specificity. Sitting alongside it are an artificial intelligence transparency statement and a stated position on the European artificial intelligence regulation, and the platform is architecturally explicit besides, exposing separate gateway services for language models, speech recognition and speech synthesis, which tells a buyer the design is model agnostic by construction rather than tied to one supplier.

Held at B rather than A on the distinction this index applies consistently: provenance is not performance. No accuracy figures, no evaluation method, no model cards, no containment or task completion rates, and no statement of which model serves which function. Naming your suppliers tells a buyer where the data goes, not how well the thing works. The separate claim of an absence of fabricated output, made in customer facing marketing, is addressed on the governance axis and is not credited here.

This is the most complete model supply chain disclosure located in the voice and conversation segment, and the comparison is not close. The trust centre publishes a named sub processor list covering the cloud platforms and the model providers behind the product, spanning two cloud infrastructure vendors, two large language model vendors and a speech synthesis vendor.

Across more than fifty voice and conversation vendors reviewed for this index, nothing else has published its model supply chain at this level of specificity, and this is the segment where the question has gone unanswered on record after record: a patient conversation passes through a speech recognition layer, a reasoning layer and a synthesis layer, and almost every vendor names none of them.

Sitting alongside it are an artificial intelligence transparency statement and a stated position on the European artificial intelligence regulation. The architecture reinforces the disclosure rather than merely accompanying it, exposing separate gateway services for language models, speech recognition and speech synthesis, which tells a buyer the design is model agnostic by construction rather than tied to one supplier, and that in turn means the published list describes a real and substitutable set of parties.

Two residuals to obtain rather than assume: retention of call audio as distinct from derived transcripts and structured data, and whether a caller is told they are speaking to software, which matters more on an inbound line than an outbound campaign.

Corrected on 5 August 2026. An earlier reading of this record stated that no named customer had been located, and that was wrong: a large multi hospital system in Texas is named publicly with a stated staffing capacity gain, and the scale figures have moved with it, now given as more than 70 million calls handled with roughly 70 percent resolved end to end without a person, alongside published figures for routine inbound automation and reduction in call abandonment.

Those numbers are large, concrete and contradictable, and a named enterprise reference is worth more than an anonymised one because the organisation has put its name to the result. What holds this at C is unchanged by the correction. Every figure is vendor produced with no baseline, no denominator, no measurement window and no independent verification.

Resolution rate is a containment measure and containment is not correctness: a call the agent resolves wrongly counts on the same side of the ledger as one it resolves well, and no accuracy, error or complaint figure is published. No peer reviewed publication and no independent evaluation was located.

Revised upward on 5 August 2026. The published trust centre answers most of what this axis asks and answers it in the place a counterparty would look: data security, access monitoring, audit logging, backups, certificates of destruction, disk encryption, a data protection impact assessment position, breach notification and a defined data subject request route.

The subprocessor disclosure is the part that matters most for a product built on models, because it tells a health system exactly whose infrastructure and whose models a patient conversation passes through, which is the single question this index has recorded as unanswered on record after record in this segment.

Two gaps keep it at B. Nothing published states whether a caller is told they are speaking to software rather than a person, which matters more on an inbound patient line than on an outbound campaign because a patient ringing their own provider did not choose to interact with a machine.

And retention of call audio as distinct from derived transcripts and structured data is not addressed, which is the ordinary unanswered question for any voice product and is sharper here given the stated volume of calls handled.

Revised upward on 5 August 2026 against a published trust centre located on a second retrieval pass. The compliance section carries an explicit health privacy rule item alongside HITRUST and an audited attestation of the operating kind, and the portal is built for exactly the counterparty due diligence a business associate relationship requires, with a documented data subject request process, a data protection impact assessment position and a breach notification statement.

That is a materially better posture than the earlier reading of this record described. Held at B rather than A because the business associate agreement itself and its terms were not located in the open, so a reader can see that the vendor is set up to enter one without seeing what it commits to.

The ownership caveat recorded earlier still stands and should be put to the vendor directly: this business has changed hands and the brand has continued over a changed product, so the contracting entity, the agreements in force and the data held under them may not be what an older document says they are.

Corrected on 5 August 2026 after a firsthand retrieval pass located a published trust centre that an earlier assessment had missed. The correction is substantial and runs in the vendor's favour. The parent company operates a hosted trust portal exposing a compliance set of HITRUST, SOC 2 Type 2, a health privacy rule position and a European data protection position, alongside a penetration test report, a cloud security alliance self assessment, and a cyber insurance statement, all available to a counterparty on request through the portal rather than described in marketing copy.

Beneath those sit published control statements across product security, data security, application security, access control, endpoint security, network security, incident response, risk management, asset management, awareness training and continuous monitoring, including audit logging, multi factor authentication, certificates of destruction, intrusion detection and automated compliance monitoring.

Two elements are rarer than the certifications and matter more for a product of this kind: a named subprocessor list and a dedicated artificial intelligence section carrying a transparency statement and a position on the European artificial intelligence regulation. This is the most complete assurance disclosure in this category and it places the record at the top of the segment scale. The honest limit is the ordinary one for a gated portal, that the reports themselves sit behind an access request rather than in the open, which is normal practice and is not a defect.

No clearance, authorisation or submission located and none needed. Automating appointment, refill and access communication is administrative. Two other bodies of law apply and neither is device regulation. Outbound automated contact falls under telephone consumer protection rules, which have been enforced against healthcare callers and which turn on the consent a patient gave to be contacted automatically.

And the company's earlier lineage in predictive outreach for care gaps and chronic disease risk touches quality measure and payment rules. A buyer checking a device framework here would be examining the wrong statute, which is now the consistent finding across every record in this segment.

Revised on 5 August 2026 to record what the artificial intelligence section of the published trust centre contains, and to add a claim that had not previously been assessed. Two things are credited. The vendor built a second artificial intelligence system to watch the first, a quality monitoring layer analysing the calls its own agents handle each month, flagging problematic interactions and allowing intervention before a pattern spreads, with its own operations lead explaining the need in terms of behaviour changing when sites launch, features ship or agent configuration changes.

That is a description of drift and of the need to monitor for it, and this index has repeatedly recorded that no drift monitoring concept exists anywhere in this market. Added now: the trust centre carries a dedicated artificial intelligence transparency statement and a published position on the European artificial intelligence regulation, which is a governance artefact almost nothing in this index offers.

Two things hold it at B. Customer facing marketing states the platform operates with no fabricated output at all. No system built on large language models can be guaranteed free of fabrication, no method exists to demonstrate it, and the claim is unfalsifiable as stated; treated here as a governance defect rather than as marketing, on the same reasoning applied elsewhere in this index, because a buyer who believes an absolute guarantee will not build the review step that would catch a failure.

And the monitoring layer is the obvious place to close the bias gap and does not: a system already reading every call could report whether interaction quality differs by the language or dialect of the caller, and nothing indicates it does, despite the product operating in several languages at scale.

Two passes located no accuracy figures, no evaluation methodology, no model cards, no containment or task completion rates, no statement of which model serves which function and no warranty, indemnity or remediation commitment. The distinction this index applies consistently decides the grade: provenance is not performance.

Naming your suppliers tells a buyer where the data goes, not how well the thing works, and this record is exemplary on the first and empty on the second, which is an unusual and instructive combination. A buyer can establish exactly whose speech recognition and whose reasoning model handles a patient call, and nothing about how often that call goes wrong. One claim made in customer facing marketing deserves naming rather than crediting.

An asserted absence of fabricated output is an absolute of a kind that cannot be verified and would be falsified by a single instance, so it functions as reassurance rather than as a measurement, and a vendor able to publish its sub processor list is well placed to publish a hallucination rate instead.

The output type raises the stakes: a patient calling their own provider receives information they will act on, and a wrong answer about an appointment, a medication or a symptom escalation is not caught by any reviewer. Ask for containment and task completion rates, the hallucination rate against a labelled sample, and what happens when the system is uncertain.

Revised upward on 5 August 2026. The parent publishes a named integration with the dominant electronic health record vendor as a first class product surface rather than as a general claim of compatibility, alongside named integrations for telephony trunking, messaging, file storage and a customer relationship platform, and a listing on a major cloud marketplace.

The agents act rather than observe: identity verification runs against the record system, availability is checked and appointments confirmed inside the same conversation, and refill requests are collected and routed into the clinical workflow.

Escalation carries structured context to the receiving human, including who the patient is, what they asked for and what has already been verified, which is an interoperability property as much as a workflow one because it requires the agent to hold state the receiving system can use.

Held at B because the depth is demonstrated for one record vendor and asserted for the rest, no second named record system integration was located, and nothing addresses write back beyond scheduling, which is where the difficult part of any record integration begins.

Revised upward on 5 August 2026 against the published trust centre. Delivery is a vendor hosted cloud service with self service sign up, and the infrastructure position is now stated rather than inferred: named cloud subprocessors, a virtual private cloud, traffic filtering, network penetration testing, backups, certificates of destruction and a documented third party risk management process.

A European data protection position with a designated representative in the European Union is published, which is a jurisdictional commitment rather than a compliance mention and is unusual for a vendor whose deployments are predominantly in the United States. Held at B because the ordinary residency detail is still missing.

No regions are named, no retention schedule is published for call audio as distinct from derived data, no tenancy model is described, and with model providers and cloud platforms on two continents in the subprocessor list, where a given patient conversation is processed is not answerable from what is published. No self hosted option is offered, which is expected at this scale and is not counted against the grade.

Revised upward on 5 August 2026. The parent publishes a pricing page and offers self service sign up with a free trial, so a prospective buyer can reach a stated commercial position without entering a sales process, which is more than most vendors in this category permit and more than the earlier reading of this record allowed. Two things keep it at C rather than higher.

The published figures were not verified firsthand in this pass and the numbers circulating in third party comparison content come from self interested sources that this index does not grade on, so what the page actually commits to should be confirmed before it is relied on.

And a self service rate card is not the same as enterprise pricing: a health system contracting at the volumes this platform describes will negotiate a bespoke agreement, and nothing published indicates how the two relate. Ownership and durability sit alongside price on this axis and remain the sharper question.

This business was acquired, its brand has continued over a changed product, and the parent has since repositioned around a general purpose platform, so a buyer should establish the contracting entity, the term, and what happens to the healthcare branded product line if the parent's centre of gravity keeps moving.

Broad across organisation type and communication channel. Customers are named as health systems, large specialty practices and care management organisations, which spans very different operating models, and the platform covers inbound and outbound contact across four channels rather than the telephone alone.

Functionally it handles the high volume access traffic of scheduling, refills and navigation, and the earlier product lineage reaches into proactive outreach for chronic disease risk and care gaps, with breast cancer and heart disease named as examples. Not tied to a clinical specialty, since patient access is not. Geography is the United States. Held at B rather than A because depth in each named organisation type is asserted rather than evidenced by a single named customer.