Phreesia

A platform that has added AI to a front office business whose value stands without it. Phreesia's core is intake: check in, appointment readiness, clinical and administrative questionnaires, consent, eligibility, time of service collection and payments, used by more than 4,700 healthcare organizations. None of that needs a model.

Phreesia VoiceAI is where the learned component sits. It answers inbound calls in the patient's preferred language, identifies the reason for the call, and resolves, routes or escalates it, and it runs outbound campaigns for reminders, payments, refills and recall; the company says 80 percent of calls are handled end to end. The pricing page also describes AI filling open appointments. Understanding a caller's intent is model work, and it is one product line among many.

The routing design is described, and the VoiceAI page promises a result no review can guarantee. The page says clinical calls get triaged correctly, every time, and its FAQ describes no more missed opportunities or errors. A promise that every clinical call is triaged correctly is a claim the output is right; a practice that believes it has no reason to audit how calls were classified, and a caller whose urgent concern was classed as routine is the failure no one sees until later.

The structure underneath is real. VoiceAI verifies the caller by name and date of birth, identifies intent, resolves routine requests such as refills, billing and records, routes urgent calls immediately to the clinical team or the on call provider through real time mobile alerts, transcribes unresolved calls to voicemail, and writes a time stamped record of every interaction to the chart. What is not published is the rule that makes a call urgent, how accurate that classification is, and whether callers are told they are speaking to an AI.

The call flow is described step by step and nothing in the technology is identified. The VoiceAI page describes answering in the patient's preferred language with no phone menu, verifying identity by name and date of birth against the practice's records, identifying the reason for the call, resolving routine requests, booking appointments, routing urgent calls, transcribing unresolved calls to voicemail, and writing a time stamped record to the chart; it says the product goes live in five days.

No model, speech recognition provider or architecture is named, no intent recognition accuracy is published, the supported languages are not listed, and nothing describes how the models are updated or how a practice learns of a change.

Every step of the call flow is described, and no model provider is named. VoiceAI's call flow, from identity check to intent recognition to routing and chart write back, is set out on its page; the security page names Bluefin for point to point encrypted payments; the company runs its own patient financing business, AccessOne; and records requests go to release of information partners, unnamed. No speech recognition, language model or telephony provider is named, and no subprocessor list is published.

The patient privacy page says data used for personalized content does not leave the platform and that sponsors receive only details a patient separately consents to share, which settles one flow and not the model chain.

Named customers with period bound figures, with the methods behind them left to the customer stories. The pricing page cites West Coast Neurology using AI to fill 44 percent of eligible open appointments, 1,600 appointments in six months and $244,000 of revenue; Effingham Health System's 55 percent increase in time of service collections in one month; a critical access hospital in Ellinwood, Kansas reaching $352,000 in digital collections; and Desert West OB/GYN's 154 percent increase in collections in one month. The VoiceAI page reports 80 percent of calls handled end to end, a 10 percent increase in appointments, a 50 percent reduction in unnecessary visits with AI assisted triage workflows and 30 percent fewer billing calls, and Metroplex Medical Centres reports 3,000 calls handled in a month. The home page states more than 4,700 organizations and one in six United States patient visits.

The VoiceAI figures carry no customer, period or denominator. Nothing measures the accuracy of call classification.

The secondary use that matters here is published in detail, with retention and training left unstated. The patient privacy page explains that after check in a patient is offered an optional HIPAA authorization; if accepted, Phreesia may use the patient's data to show personalized health content funded by life sciences companies, advocacy groups, public health bodies and insurers. Declining ends the session with no effect on the visit, the authorization can be revoked by email, the data does not leave the platform, no third party cookies or trackers are placed in content, and a sponsor receives only the limited details a patient separately consents to share. The Platform Privacy Policy says data the company holds as a business associate is governed by HIPAA and its agreements.

What is missing is the retention schedule for intake data and call recordings, and any statement on whether patient data or VoiceAI calls are used to train models. Nor does the page say how content is matched to a patient who has authorized it.

The company states it acts as a business associate, backed by a substantive privacy document, with the agreement unpublished. The Platform Privacy Policy says that when Phreesia receives health information from or for healthcare providers it acts as their business associate and enters into business associate agreements requiring it to safeguard patients' protected health information, and the patient FAQ says the same. The patient privacy page sets out the separate path by which a patient can authorize use of their data for personalized health content, and the security page lists HITRUST r2 certification of the intake platform and data centers.

The agreement itself is not published, so its tier, subcontractor terms and treatment of VoiceAI call recordings cannot be read before contracting. How the authorized content program sits alongside it is not explained.

Several recognized attestations named with their tiers, with the core reports available on request rather than open. The company's security page lists HITRUST CSF certification (the company announced r2, the risk based two year assessment, for its intake platform and data centers, and the page now dates the certification 2025), a completed SOC 2 Type 2 examination, PCI DSS Level 1 service provider status and a PCI validated point to point encryption solution listed under Bluefin, and it publishes a responsible vulnerability disclosure policy. The payment listings can be checked in public registries.

The gap is access. The SOC 2 and HITRUST reports, which are the ones covering patient data, are obtained by emailing the security team, and the page does not state whether VoiceAI, launched in 2025, sits inside the certified scope.

VoiceAI is marketed on triaging patients' clinical calls, and no regulatory position is stated for it. The VoiceAI page says the agent listens, understands the patient's concern and routes urgent issues to the right clinical team, that "clinical calls get triaged correctly, every time", and that practices see a "50% reduction in unnecessary visits with AI-assisted structured triage workflows". The agent decides from a caller's description whether a concern is urgent enough for the on call provider, and the visit figure shows that the same sort changes whether a patient is seen at all. The page says next steps follow the organization's established workflows; it names no published triage protocol and no nurse or clinician making the disposition, and presents the product as cutting nurse line overhead. What happens to a call the agent marks routine is not described.

The rest of the platform is administrative: intake, consent, eligibility, payments and scheduling carry no clinical claim. The company is also subject to securities disclosure, and investor class actions filed in May 2026 concern its fiscal 2027 guidance rather than its products.

Nothing published on how VoiceAI's behavior is evaluated or monitored. The company markets answering calls in the patient's preferred language and classifying clinical concerns, and publishes no accuracy by language, accent or caller group, no evaluation method and no drift monitoring. A caller the system misunderstands at the front door is the one who does not reach the right clinician or appointment, and multilingual operation marketed as a strength with no per language results is the central exposure.

The question the content program raises is now answered in part. The patient privacy page says content is personalized only after a patient accepts an optional HIPAA authorization, that data does not leave the platform, and that clinicians see in aggregate what content patients may see; it does not say how content is selected for a patient or how that selection is reviewed.

Every call leaves a record that can be checked, and nothing stands behind the outcome. VoiceAI writes a time stamped audit trail of every interaction to the chart, transcribes unresolved calls to voicemail, and alerts the on call provider in real time for urgent calls with two way texting on the record. A practice can therefore see what the system did with each call.

No accuracy figure for intent or urgency classification is published, and the site carries no warranty, service level on call handling or remediation commitment; the contract terms could add one. The product markets correct triage every time; nothing says what the company answers for when a call classed as routine was urgent, or whether a practice can sample the calls not marked urgent.

Two way integration with record systems is claimed and partly verifiable. The pricing page says Phreesia integrates bidirectionally with many of the leading record systems, and the VoiceAI page says every call is written back to the chart and appointments are booked into the calendar. The company is listed in Veradigm's partner marketplace, and an Epic listing has been reported but not confirmed.

No record system is named on the pricing or VoiceAI pages, a separate integrations page may list them, and no interface standard or write back method is published.

Nothing published says where the platform runs or where data rests. The company's HITRUST announcement refers to its intake platform and data centers, which shows the company operates hosting environments, and no page names a provider, region, residency commitment, tenancy model or retention period. The product spans waiting room devices, patients' own phones and the telephone, so intake answers, card payments and call recordings are three kinds of data with three residency questions.

The pricing mechanism and the current offer are published; a price is not. The pricing page says pricing is customized by organization size, workflows and the products used, as a predictable subscription with no per message charges and no fees, and that the full Phreesia Intake platform is free through 31 December 2026 before standard pricing begins. Most practices go live in about six weeks, and VoiceAI is described as live in five days. As a listed company it reports revenue per client in its filings, which lets an outside reader estimate average spend.

What no page states is a rate, how VoiceAI is priced, the payment processing rates, or how the payments business and the sponsored content program bear on a practice's terms.

Very broad by reach and confined to the front of the visit. The company states it serves more than 4,700 healthcare organizations and one in six United States patient visits, with pages for health systems, enterprise groups, independent groups and specialties. Named customer stories cover neurology (West Coast Neurology), obstetrics and gynecology (Desert West OB/GYN), a critical access hospital in Kansas and a health system in Georgia, and VoiceAI quotes come from Metroplex Medical Centres and Evergreen Family Medicine.

Functional coverage runs from scheduling and intake through eligibility, consent, payment, patient financing and now calls. It stops where care begins, and the VoiceAI results are not tied to named settings beyond two quotes. United States only.