Parlance

Speech recognition is the mechanism the whole product rests on. A caller speaks a department, a clinician name or a request, and the system understands it and acts. Remove the models and there is no product, only a directory. The reason this sits at B rather than A is that the company says so itself, unusually plainly: it argues in its own published commentary that every vendor in this category runs on essentially the same core technologies from the large cloud and model providers, and that the real differentiator is not the underlying model but the proven capability to implement and integrate it and to perform.

On that account a substantial share of what a health system buys here is the managed service, the directory maintenance, the routing design and thirty years of deployment method rather than a proprietary model. That is a defensible commercial position and it is also an accurate description of a vendor whose artificial intelligence is central to the product but not proprietary to the company.

Automation is high and the handoff is designed rather than bolted on. Published figures put routine call automation above 85 percent, agent time saved above 40 percent, and a drop in live agent call volume above 33 percent on the first day, with HCA Healthcare described as handling more than 70 percent of calls automatically across more than 75 hospitals. Two design choices earn the grade.

Complex, sensitive, billing and care coordination calls are explicitly listed as the human team's work rather than automation targets, which is a rarer piece of restraint than it sounds in this segment. And an escalated call arrives at an agent with context already populated, so the patient does not begin again.

The fully managed service model is itself an oversight structure of a kind, since the vendor monitors and tunes the system continuously against agreed service levels while the health system directs escalation rules and scheduling guidelines. Held at B on two unpublished points.

The Parlance 12 semantic routing feature deliberately moves an unrecognised caller intent to a large language model instead of straight to a live operator, which is an upgrade that reduces the human fallback, and nothing published describes what bounds that model or what happens when it is also wrong. And a health system switchboard receives emergencies in the same queue as parking questions, with no published criteria for how the system recognises that moment or what it does next.

An unusual split, candid at the category level and quiet about itself. The company states in published commentary that every vendor in this space operates on essentially the same core technologies from Google, Amazon and other major providers, and that the differentiator is implementation rather than the model.

Across more than fifty voice and conversation vendors reviewed for this index, almost none says anything about model provenance at all, and this is the most explicit statement located that the models underneath the category are bought rather than built. It does not follow that a buyer learns what they need.

Which provider serves this platform, which model handles the semantic routing introduced in July 2026, whether inference happens inside the vendor's environment or at a third party, and whether the answer differs for a customer in the United Kingdom are all unanswered, and those are precisely the questions the candid statement raises. Named components without named models: semantic routing to a large language model for unresolved intent, augmented recognition, synthesised voice. No model cards, no accuracy figures, no evaluation method.

One published statement lifts this above the vendors that say nothing, and one feature nobody would think to look at is the substantive finding. The security policy makes a product level data statement rather than a website one: customer data is encrypted at rest and transmission is always encrypted in transit.

Against that, the privacy policy explicitly limits itself to the company's websites and states in its own text that it does not apply to information collected by other means, so the only published privacy document deliberately excludes the product that handles patients. Nothing addresses whether calls are recorded, how long audio is retained, whether recordings improve models, or whether a caller is told they are speaking to software. The finding is the directory routing.

Patient room routing runs on feeds from the record system, which means the system can connect a caller who names a patient to that patient's room, and confirming that a named person is an inpatient is a disclosure decision governed by directory opt out rules rather than a routing convenience. That decision is being made here inside a routing function, by a machine, at conversational speed, for every caller who names a name.

How the opt out flag propagates from the record system into the routing layer, and what the system says when a patient has opted out, is the question to put in writing, because the failure mode is a confirmation the patient specifically asked not to be given. Ask that, plus recording, retention and the sub processor list.

The operational denominator here is the largest in this category by a wide margin and it is concrete enough to be contradicted: more than 1.9 billion patient calls answered, a sustained rate above six calls per second, more than 400 health system brands.

The customer evidence is named rather than anonymised, which matters, and several figures are specific: Virtua Health doubling annual call volume from 200,000 to 400,000 without adding contact centre staff, UW Medicine at close to 4 million calls a year, a New Jersey health system publishing a saving near one million dollars. What holds this at C is the shape of the evidence rather than its quantity.

The headline percentages are floors expressed with a plus sign, published by the vendor, with no baseline, no denominator, no measurement window and no method. Scale of use is not evidence of benefit, and a containment rate is not the same as a correct outcome, because a call the system contains but resolves wrongly counts on the same side of the ledger as one it resolves well.

No peer reviewed publication and no independent evaluation was located, and the metrics the company says it commits to before a contract, containment, first call resolution, agent hours saved and abandonment, are reported to the individual customer rather than published.

The published security policy does make a product level data statement, which lifts this above the vendors that say nothing: customer data is encrypted at rest and transmission is always encrypted in transit. Three gaps sit against it.

The privacy policy that exists explicitly limits itself to the company's websites and states in its own text that it does not apply to information collected by other means, so the only published privacy document deliberately excludes the product that handles patients. Nothing published addresses whether calls are recorded, how long audio is retained, whether recordings or transcripts improve models, or whether a caller is told they are speaking to software rather than a person.

The sharpest point is a feature nobody would think to look at. Patient room directory routing runs on HL7 feeds from the record system, which means the system can connect a caller who names a patient to that patient's room. Confirming that a named person is an inpatient is a disclosure decision governed by directory opt out rules, and it is being made here inside a routing function by a machine at conversational speed. How the opt out flag propagates from the record system into the routing layer, and what the system says when it does not, is the question to put in writing.

The company describes its voice assistants, its text messaging and its web chat as compliant with the federal health privacy rule consistently across every product page, and displays a compliance badge in its site footer. That is a clear assertion of position and it is made repeatedly rather than in passing. It remains an assertion.

No body certifies compliance with that rule, so a badge records an intention rather than an audited finding, and the documents that would let a counterparty check the position were not located in two passes: no business associate agreement terms, no compliance page, no notice describing the vendor's role.

The role itself is not in doubt, since a supplier hosting and operating a health system's patient telephone line is plainly a business associate, but the terms on which it holds that role are unpublished. Given the July 2026 release routes unrecognised caller speech to a large language model, the specific term worth reading is which subprocessors are permitted and on what basis.

A published security policy that describes specific controls rather than gesturing at a posture, which is a different and in some ways more useful disclosure than the badge most vendors offer. It names access on least privilege, encryption of customer data at rest, transport encryption using a stated protocol version or a network layer tunnel, third party external penetration testing at least annually with findings remediated inside a service level, internal vulnerability scanning, centrally managed and mobile device managed endpoints with antivirus, virtual private network access for remote connectivity, and mandatory monthly staff security training.

An annual independent penetration test is a genuine outside check, and committing to a cadence for it is more than most publish. What is absent is an attestation. No report against a recognised control framework was located, no information security management certification, no health sector assurance credential, no trust centre and no artefact a procurement team can request under a non disclosure agreement.

The distinction that decides the grade: a control description tells a buyer what the vendor says it does, while an audited report tells them an independent party tested whether it actually happens. Placed against the other conversation vendors in this index, this sits above those that publish nothing and below those that publish a named audited report of the operating kind.

No device clearance, authorisation or submission was located and none would be expected. Routing a call, booking an appointment and reporting prescription status are administrative functions, and the product does not offer clinical assessment or advice. The regulators that do apply are elsewhere and the company operates on the edge of two of them.

Outbound automated calling and text messaging carry telephone consumer protection obligations that have been enforced against healthcare callers, and the platform runs outbound as well as inbound. Operating in the United Kingdom brings that deployment inside a separate data protection regime with a common law duty of confidentiality alongside it.

The wider gap is the one this index has recorded across the segment rather than a failing of this vendor: there is no regulatory standard at all for artificial intelligence occupying the first contact position on a health system telephone line, which is where a person describing an emergency arrives. A switchboard answering at more than six calls per second occupies that position more often than any other product in this index.

This is the largest instance of the speech recognition equity gap the index has recorded, measured by exposure rather than by conduct. Recognition accuracy varies with accent, dialect, first language, age, hearing impairment and emotional state, all of which are concentrated in exactly the populations that call a hospital switchboard.

The company markets recognition quality as a core strength and the July 2026 release names a significant improvement in first pass recognition accuracy, so accuracy is measured internally. None of it is published, and none of it is broken down by caller population. Against 1.9 billion answered calls across 400 health systems in three countries, a difference of a few points between caller groups is a very large number of people bounced to a queue, misrouted or misunderstood.

The failure mode is what makes it a D rather than a C: a caller the system did not understand produces no incident, no complaint and no record of why, so the organisation running it cannot see the disparity even in principle. No published position on bias assessment, on model or vendor selection criteria, on drift monitoring after a configuration change, or on human review of routing decisions was located. A vendor operating a managed service and reviewing interaction data continuously is better placed than almost anyone in this index to report this, which is what makes the silence notable.

An unusual split sits on this record: candid at the category level and quiet about itself. The company states in published commentary that every vendor in this space operates on essentially the same core technologies from a small number of major providers, and that the differentiator is implementation rather than the model.

Across more than fifty voice and conversation vendors reviewed for this index, almost none says anything about model provenance at all, and this is the most explicit statement located that the models underneath the category are bought rather than built.

It is worth carrying beyond this record, because it reframes what a buyer should be comparing: if the base models are common, then claims of superior model quality are mostly noise and the questions that matter are grounding, escalation design, integration and evaluation. Naming that is a service to buyers even though it is not a disclosure about this product. It does not follow that a buyer learns what they need here.

Which provider serves this platform, which model handles the semantic routing introduced in 2026, whether inference happens inside the vendor's environment or at a third party, and whether the answer differs by jurisdiction are all unanswered, and those are precisely the questions the candid statement raises. Named components without named models, no model cards, no accuracy figures and no evaluation method. Ask for containment, transfer rate and the routing model.

Integration is load bearing here rather than decorative, which is what separates this from a telephone product with a healthcare label. The platform reads live provider schedules so appointment availability is current at the moment a caller asks, writes bookings and changes back, surfaces prescription and referral status, populates the agent desktop before a transferred call arrives, and refreshes directory data automatically so routing does not decay as staff move.

Patient room routing runs on HL7 feeds. The company states it integrates with all major record systems and that its customer solutions team includes people who came from record system vendors, which is a credible claim given deployments at organisations running each of the dominant platforms. Held at B because the claim is general where it should be specific.

No record system is named, no marketplace listing or partner status was located, no position on the modern interoperability standard is published, and all major systems is an assertion a buyer cannot check against a list. For a health system, the question that decides the project is not whether integration exists but which of its own instances have been done before and how long the last one took.

A single delivery model, described plainly. The vendor hosts, monitors, maintains, tunes and updates the platform, publishes service levels and escalation procedures, and holds the technical work so the customer's team needs no voice or artificial intelligence expertise. The customer portal is a vendor operated cloud service.

Integration with the incumbent telephone estate is a genuine strength for a switchboard replacement, with the platform sitting alongside the major enterprise telephony systems rather than requiring their removal. No self hosted or in place option is offered or would be expected at this size. Residency is the gap and operating in three countries makes it a real one rather than a formality.

Nothing published states where call audio, transcripts or derived data are processed and stored, whether a customer in the United Kingdom or Canada is served from infrastructure in their own jurisdiction, which cloud regions are in use, or how long anything is kept. A United Kingdom health organisation has a specific reason to ask, because moving patient voice data outside that jurisdiction engages transfer rules that a United States buyer never has to think about.

No pricing, no rate card and no basis of charge is published, which is ordinary for this segment. The durability question resolves differently and better than usual. Parlance was acquired by Volaris Group in 2024 and is a wholly owned subsidiary of Constellation Software, an acquirer listed in Toronto that operates a publicly stated buy and hold strategy and does not customarily divest the businesses it acquires.

For a health system placing its main telephone line with a supplier, that is a materially different continuity position from a venture backed vendor working through a funding runway, and the company argues the point openly in its own material. A distinction worth holding onto, though, because it is easy to read too much into: being owned by a listed company is not the same as being one.

Constellation reports at group and operating group level across more than 1,100 businesses and does not publish results for a business of this size, so the group listing gives a buyer confidence about survival without giving them a single figure about this supplier. Third party estimates of revenue exist but vary widely and none is authoritative. The company does publish a stated time to return of about 30 days and names the operational metrics it will be held to, which is more commitment than most.

Breadth of organisation type is the strong half. More than 400 health system brands spanning a national hospital operator with more than 75 sites, academic medical centres, a children's hospital, community and regional systems, and specialty and multi site clinics, across the United States, Canada and the United Kingdom.

Because the product sits on the switchboard, it touches every department, every clinic and every inpatient unit in an organisation by default rather than by design, and it reaches patients through voice, text message and web chat. Two things hold it at B. Language coverage is not published at all, which is a real gap for a product answering the front door of safety net and urban health systems, and it is the same gap that appears on this record's governance axis.

And the buyer is a provider organisation in every case located, with no payer, pharmaceutical or public health deployment, so the span is deep within one segment rather than across the industry. Worth recording for scope: the company also sells to enterprise and higher education, which sit outside this index, and the higher education line is a legacy vertical rather than a healthcare package.