TalkToMedi

The agent is the product. MEDI answers inbound calls in parallel, books and reschedules against each clinic's schedule, captures intake, turns voicemail into structured requests that continue by text, and places outbound calls and texts for referrals, recalls and preparation reminders, all under rules the clinic sets. A cofounder's newspaper column describes voice technology the company built for clinical accuracy, and the website sample call runs on OpenAI's voice service.

Take the models away and nothing remains to buy. The clinic keeps its own record system and phone line, and the company sells no scheduling or record software of its own.

Each clinic writes the rules the agent works within: workflows, approved answers, permitted actions and where calls escalate. Urgent, complex, uncertain or patient requested handoffs go to the responsible staff member with the call's context, a caller describing an emergency is told to hang up and call 911, and setup tests ordinary calls, edge cases and a transfer nobody answers before launch. Outbound referral calls start only from referral tasks staff have already reviewed in the record system.

The company states that 72 percent of calls are resolved end to end, and no review is described for the routine calls the agent closes on its own. Telling callers they are speaking with an AI is left to each clinic under the terms, which say the company may add standard disclosures to a workflow.

The production stack is not described. OpenAI is named as the voice provider for the website sample call and appears among the company's supporters and ecosystem partners, while a cofounder's column says the company developed its own voice AI around accuracy and clinical knowledge. Which of these answers a patient's call, and how the parts fit together, is not stated.

There is no architecture description, base model, accuracy figure or release history. The comparison page rates the company's own specialty intake as beta and leaves empathy and complex judgment to staff.

One provider is named, for the demonstration only. The privacy policy says the website sample call streams microphone audio straight to OpenAI with a request not to retain it, and the terms list telephony, hosting, speech recognition and AI model providers by category. No provider is named for patient calls, and no subprocessor list is published.

A patient call passes through telephony, speech recognition, a model and voice synthesis before anything is booked. The privacy policy binds service providers to protect the data by contract.

Named clinics report operational results. At Dr. Neil Ingber's medical clinic in Scarborough, Ontario, the agent handles more than half of call volume, 65 percent in an updated figure, with shorter waits and fewer dropped calls; Hunter Medical reports an 89 percent fall in hold times. One unnamed clinic logged 5,666 bookings, 18,979 resolved calls and 5,206 after hours calls over 95 active days.

None of it is independently measured. The platform figures (72 percent of calls resolved end to end, more than 100 hours of front desk time a month, more than $70,000 a year per clinic) carry no method or denominator, and nothing reports how often the agent misheard a caller or misjudged urgency.

Retention is stated in numbers: call audio for up to 48 hours, transcripts and summaries for up to 30 days, and deidentified metadata kept for service improvement. The privacy policy says personal information is never used to train foundation models and that voice recordings are not used for model training.

The terms are wider than the policy. They bar training a general purpose or foundation model on identifiable patient data, audio or transcripts without the clinic's written permission, and allow deidentified data, evaluation data and logs to improve the service; the deidentification method is not described. For safety engineering, the company describes test cases drawn from real edge scenarios, evaluation on unclear and emotionally charged calls, and specialty escalation rules developed with clinicians at the Peter Munk Cardiac Centre through UHN's ECHO program. No results from that testing are published.

A business associate agreement is named in the terms as one of the documents a clinic can sign, and the privacy policy reaches the service. The company acts as service provider and processor, practitioners remain custodians of patient records, and data crossing the border carries safeguards described as consistent with PHIPA, PIPEDA and HIPAA. No agreement or template is published.

The company's pages pull in two directions on scope. The home page shows a HIPAA mark captioned privacy and security, while the trust page describes HIPAA aligned controls whose scope each clinic is told to confirm. On the Canadian side, completed privacy impact assessments for British Columbia, Alberta and Ontario are available on request and are described as the company's own materials, not regulator certifications. The privacy policy lists a Gmail address as its privacy contact.

No security attestation has been issued. The trust page says a SOC 2 Type II program is underway and states that it presents no issued report, and the home page shows a SOC 2 Type II mark captioned compliance underway. A November 2025 post says outside compliance review firms, Delve among them, validate the company's controls, and lists encryption in transit and at rest, role based access, audit logging and documented incident response.

No trust portal, penetration test summary, report or certificate is published, so each control rests on the company's own description. The platform holds call audio, transcripts and record system credentials for clinics in seven provinces.

Triage is marketed with no regulatory position stated. The privacy policy lists basic triage and routing among the services, the terms list configurable triage and routing workflows, and a November 2025 post describes specialty triage logic that sends chest pain, shortness of breath and symptoms after a procedure down dedicated escalation rules. Callers describing an emergency are told to call 911, and with 72 percent of calls stated to close without staff, the agent decides which calls are routine.

The terms say the service does not determine whether a matter is urgent and leave triage, escalation and emergency instructions with the clinic, which allocates responsibility rather than stating a position. No Health Canada or FDA position is published, and no named triage protocol or clinician decision on every sorted call is described.

Testing is described, and no evaluation is published. The company says it builds test cases from real patient edge scenarios, evaluates behavior on unclear, partial and emotionally charged calls, keeps transcripts and summaries for quality review, tests each clinic's workflows before launch, and drew on clinicians at the Peter Munk Cardiac Centre for cardiology escalation. Nothing reports how often the agent understood the caller, booked the right visit or caught an urgent call.

The agent listens to patients across seven provinces, in specialties from pediatrics to neurology, and its own comparison table rates it multilingual. Speech recognition varies with accent, first language, age and hearing, and no result is broken down by any of them, so errors could gather among the callers who sound least like the data the system learned from.

The terms say plainly that outputs can be wrong and place the consequences on the clinic. They warrant only commercially reasonable skill and care, disclaim the accuracy of AI outputs, state that the service may miss urgent conditions and transcription errors, cap each party's liability at 12 months of fees, and have the clinic indemnify the company for claims arising from triage, emergency response and patient care. Clinics must apply qualified human review before relying on outputs for consequential actions.

Handoffs carry the call's context to staff, and transcripts and summaries are kept for quality review, so a clinic can find errors after the fact. No error rate is published, and a patient has no route to correct a call outcome except through the clinic.

Systems are named and booking into them is shown, with depth varying by system. The integrations page lists Juvonno, OSCAR Pro, Accuro, LibreMD and Epic as partnership integrations, and TELUS Health CHR, Med Access and Pomelo, eClinicalWorks, Open Dental, Dentrix, Noterro and iMedicWare as agent connectors; the Juvonno entry links that company's partner directory. The home page shows bookings matched by provider, visit type and availability, a pain clinic intake saved to the record, and outreach worked from referral tasks in the record system.

What a partnership integration covers, how an agent connector differs from one, and which systems accept writes are not explained. Available actions depend on each platform and its configuration, and the company's comparison table rates appointment management in the clinic's system as requiring setup.

Canadian residency is described, then qualified. A November 2025 post says all patient communication data is stored and handled on Canadian servers, and the trust page refers to Canadian infrastructure. The privacy policy, updated in October 2026, says Canadian clinic data is stored in Canada where possible and that some account and billing information is processed in the United States, and the company's own comparison table marks its Canadian data residency for review.

One hosted service sits over each clinic's record system and phone line. No hosting provider or region is named, the telephony, speech and model services a call passes through are not identified, and data handling for clinics outside Canada is not described.

The contract terms are public and the price is not. The user terms (June 2026) place fees, implementation charges, renewals and overages in each order form and in a support and payment terms document the site does not link, renew subscriptions for the same term unless notice comes 30 days before it ends, and make fees non refundable unless an order form says otherwise. A self serve signup sits beside the demo route with no rate, unit or tier shown before it.

Only the value side is quantified: more than $70,000 a year of operational value per clinic, and a sample outreach campaign worth $1,440 at $120 a booking.

Coverage is named in detail, while results come from general medical clinics. The company publishes pages for family medicine, clinic networks, pain, sports and physiotherapy, allied health, cardiology, pediatrics, neurology, endoscopy, dermatology and diagnostic imaging, each setting out its booking and preparation rules, and says it supports clinics in seven provinces caring for more than 1.5 million patients. Dental systems appear among its integrations, and the Epic entry mentions hospital and health system workflows.

The named results come from Ontario medical clinics, and no hospital deployment is named. The company's comparison table marks its own specialty intake and workflows as beta, so the specialty pages describe scope more than demonstrated use.