Athel AI

Every module runs on a model. The intake interview is conducted by the AI itself, which chooses each follow up question from the patient's last answer and turns the conversation into a structured history with items flagged for review. Reception is an AI agent that answers the clinic's phone and texts, classifies each call and books visits. Dictation converts speech to text as the clinician talks. Take the models out and there is no interview, no receptionist and no draft left to sell.

The intake is built as several narrow agents rather than one model: one conducts the interview, one structures the output and one checks for red flag signals, so a failure can be traced to the stage that produced it.

Intake hands off rather than decides. Every summary is staged for a clinician to review before it reaches the chart and is never filed automatically, red flag findings are surfaced for review, and any risk stratification is rederived by a separate check before the report is released. The terms require the clinic to review and approve every output and state that the service makes no autonomous clinical decisions.

The gap sits on the phone. Reception answers every call around the clock and sorts calls into urgent, medical, refill, scheduling and admin queues on a staff dashboard, but nothing describes what happens when a caller is in crisis: whether the call transfers to a person, what the caller hears, or how an urgent call is handled after hours. The marketing also promises that no patient detail is ever lost in translation, which a physician reviewing the summary has no way to check.

The design is described and the models are not named. The intake runs as several specialized agents, one conducting the interview, one structuring the output and one checking for red flags, and the terms say the service runs on large language models and speech processing systems. The interview follows a three cycle framework: the presenting symptom, associated symptoms, then relevant history. The partners page says the models are fine tuned for structured intake, pre assessment and triage.

No model, provider, version or update practice is identified, and no accuracy figure exists for the interview or the summaries, so a buyer cannot tell when the system changed or what changed.

The roles in the chain are named and the companies filling them are not. The privacy policy lists a cloud infrastructure provider, a managed database provider, voice and telephony providers and a payment processor, and the terms describe AI infrastructure providers bound by contract not to train on patient data. Identifiers are stripped before any text reaches a model, and providers handling patient data run in the customer's own country.

The subprocessor list sits in a trust center shared only under NDA. The partners page shows Google, Microsoft Azure and OpenAI in a strip of organizations the company works with, alongside SickKids and the University of Toronto. That is a logo display rather than a statement of which model processes patient data.

The figures come from the company's own deployments, with no clinic named and no method behind them. A specialist clinic case study reports 3.6 times more clinical data captured in half the time, with question completeness rising from 27 percent to 96 percent. A primary care example reports 10 minutes saved per appointment. The hospital triage page states that 97 percent of patients who start an intake finish it, and running counters on the home page report 6,850 in clinic encounters, 2,137 physician hours saved and 51,557 clinical calls automated.

None of these figures carries a denominator, period, site count or comparison, and no customer is named. No peer reviewed study or independent evaluation exists for the interview or the receptionist. The company reports winning the Clinical Excellence Award at BVX 2026, which recognizes the product without measuring it.

Training, de identification and retention are each stated specifically enough to hold the company to. The terms of service say customer data containing personal or health information is never used to train, fine tune or improve any model, the company's own or a provider's, enforced by contract with its AI infrastructure providers and by removing identifiers before any model sees the text. Names, birth dates, health card and phone numbers, and clinic and physician identifiers are stripped, so the model receives only the clinical content of the conversation.

Records are kept seven years from the last entry by default, adjustable by contract, with a 30 day export window after termination, deletion by destroying the encryption keys, and backups purged within 90 days. A company article gives a different figure for call audio and transcripts: at least 10 years.

On safety engineering, every report passes a grounding check against the transcript and an independent rederivation of any risk stratification before release. How a safety event is handled once found is not described.

Both the terms of service and the privacy policy state that the company acts as a business associate for US customers. A written agreement must be signed before any protected health information reaches the company, the agreement controls wherever it conflicts with the terms, and subcontractors handling that information must sign agreements carrying the same obligations. Breach notice to the covered entity is set at no later than 60 days in the privacy policy, while a company article on Canadian privacy describes a 24 hour contractual window.

The Canadian side is set out to the same depth: the company acts as an agent of the health information custodian under Ontario's PHIPA and as a service provider under PIPEDA. The agreement itself is shared only during contracting, so its scope and terms cannot be read in advance.

Controls are described in detail and nothing independent stands behind them yet. The terms say the company is actively pursuing SOC 2 Type II, so no attestation is held today, and no penetration testing cadence is stated. The privacy policy lists AES-256 encryption at rest, TLS in transit, encrypted voice streams, per clinic isolation, role based access with multifactor authentication, an append only audit log of every access to patient data, anomaly alerts and a tested incident response plan.

A trust center holding the subprocessor list, a security white paper and any attestations is open to customers and prospects only under NDA, on request. The HIPAA, PIPEDA and PHIPA badges across the site are compliance claims rather than audits.

The regulatory position is stated on every page: the company says its applicable functions are intended to qualify as non device clinical decision support under US and Canadian frameworks, that it does not diagnose or treat, and that clinical judgment stays with the clinician. The terms add that the service is not a diagnostic device and that its status is assessed under Health Canada's Medical Devices Regulations and the FDA framework for software as a medical device. The hospital triage page matches that position, saying the product records what patients report and does not assign triage levels or prioritize patients.

Other pages pull against it. Reception is described as triaging every patient and ranking calls by priority, every intake report carries urgency and risk flags, and the urgent care protocol is described as automatic risk stratification with escalation flags for emergencies. No clearance or registration is held, and the non device position is the company's own reading rather than one a buyer can look up.

Language is where governance matters most for this product, and nothing is shared on it. The company markets interviews in any language and says no patient detail is ever lost in translation, but lists no supported languages and gives no completion or accuracy result for any of them. A patient the interview misunderstood leaves no complaint, and the physician reads a summary without seeing what was lost.

The commitments that do exist concern process. Patients are told they are speaking with an AI before the interview starts, no decision with significant effects is made by automated processing alone, and the system is designed as narrow agents whose failures can be traced to a stage. No fairness statement, subgroup result or evaluation method is available.

Clinician review is the safeguard, and the terms leave the clinic carrying the output. The service is provided as is, AI outputs are described as probabilistic with no guarantee of accuracy, completeness or clinical validity, and liability is capped at twelve months of fees. The company's indemnity covers intellectual property claims only, while the customer indemnifies the company for any clinical decision made in reliance on the service.

The means to catch an error are real. The clinician reviews each report before it reaches the chart, a grounding check ties every clinical claim to the transcript, and transcripts and audit logs stay with the clinic. Patients correct their records through the clinic, which the company supports. The 99.9 percent uptime and 100 percent call answer guarantees on the pricing page cover availability, not what the output says, and no error rate is shared.

Clinic plans write into named record systems in one direction. After the physician reviews it, the structured report moves into the chart in Accuro, OSCAR or TELUS, with Accuro connected through a secure workflow extension, and clinics can authorize their EMR to send demographic and appointment data to Athel. Epic, Cerner and Meditech connect through HL7 and FHIR on the Enterprise tier, and the partners page describes two way sync with TELUS PS Suite, OSCAR, Epic and Cerner for white label partners.

No marketplace listing, EMR partner designation or named site running any of these connections is available. Around the visit, Reception syncs with the clinic's calendar to book, move and cancel appointments, and the partners page names TimeTap and Calendly among supported booking platforms.

Residency is committed rather than implied. Canadian customers' data is stored and processed in Canada and US customers' data in the United States, including transcripts, audio, AI inference, voice synthesis and backups, and patient information does not cross the border. The only processing that does cross is operational data with no health information, such as support tickets, billing records and analytics events, under data processing agreements.

Each clinic's records are isolated, described in the privacy policy as row level security at the application and database layers and in a company article as dedicated database instances. The service runs on enterprise cloud infrastructure whose provider is not named. Enterprise customers can white label the patient experience on their own domain and embed it through an iframe, SDKs or an API.

Prices are listed in full for three self serve plans, each charged per physician per month: Basic at $149, Pro at $399 and Premium at $699, with annual billing 10 percent lower. Verified physicians can start a 30 day free trial without a sales call. Each plan sets monthly volumes per physician, from 350 calls and 500 intakes on Basic to unlimited calls and 2,500 intakes on Premium, and usage beyond those limits is billed on a scaled basis with no rate given.

Enterprise pricing for hospitals and multi site groups is quoted, with a 30 day pilot on approval. The pricing page does not say whether figures are in US or Canadian dollars, which matters for a company selling on both sides of the border. The terms let fees rise at renewal on 30 days' notice and give no refunds for partial periods.

Coverage is named in detail. Specialty pages cover psychiatry, cardiology, gastroenterology, oncology, family medicine and walk in clinics, and hospital triage, where patients complete the interview on their own phone in the waiting room. Interview protocols are described for primary care, cardiology, orthopedics, dermatology, mental health and urgent care. Interviews run by voice, chat or video, pediatric interviews are conducted with a parent or guardian, and the markets are Canada and the United States.

What is missing is evidence tied to a named setting. The case studies describe an unnamed specialist clinic and unnamed primary care practices, and no deployment is named in any specialty.

Athel also appears among healthcare administrative automation vendors because Reception handles patient access work: booking, moving and canceling visits, a patient booking portal, reminders, and health card capture at registration.