Talkie.ai

The AI agents are the product. Each one holds the conversation by phone, web chat or text, works out what the patient needs and completes it in the practice's systems: booking, rescheduling or canceling a visit in the EHR, collecting refill details, registering a new patient, answering questions about hours, directions or visit preparation, and filing a message for the care team when a person has to act.

Outbound work runs on the same agents, which call patients for confirmations, waitlist openings and recalls the EHR shows as due. Remove the language and speech models and nothing remains to sell; the implementation team, phone integrations and EHR connections exist to put the agents to work on a practice's real call volume.

Practices set the transfer rules, and any caller can ask for a person at any point; patients are told they are speaking with an AI assistant. During office hours the agent hands calls it cannot resolve to the front desk. After hours, with no one to forward to, it resolves routine requests itself and routes true medical emergencies to the on call provider by transfer, pager or text, as each practice configures. Messages and cases land in the EHR for staff to work the next morning.

The criterion is the open piece. Nothing defines what the agent treats as a real emergency, what it does when a caller describes worrying symptoms without asking for the on call provider, or how a wrong routing decision is caught. Talkie advertises that every call is picked up around the clock, which makes the after hours rule the control that matters most.

The language, speech recognition and telephony models behind the agents are not identified. The architecture is described only in outline: AI capabilities sit on model, speech and telephony subprocessors, only plain speech to text transcription is allowed, and the data flow through those parties is mapped for each client.

The change discipline is stronger than the disclosure. Prompts, models and provider settings are version controlled and released through the same review as other software, production calls emit traces recording the model and provider used, and agent performance is reviewed against a baseline on a set cadence. A practice learns which models handle its calls through its own data flow mapping after signing, not from anything shown before buying.

Amazon Web Services is the named host, and the language, speech recognition and telephony providers behind the agents go unnamed. The terms governing those parties are stated: each AI provider must hold a business associate agreement or data processing agreement matching the data it touches, every subprocessor that may process patient information signs a business associate agreement, and AI providers are contractually barred from training on Talkie or customer data.

Where patient information enters and leaves that chain is mapped for each client, which gives every customer its own record. No public subprocessor list exists, so a prospective buyer cannot see before signing which outside model or speech company hears its patients' calls.

Named practices report call volumes and staff relief, with no method or baseline behind the savings figures. ESD Pediatric Group, a two location Cincinnati practice live on athenaOne since April 2026, says its agent handles more than 6,000 calls a month and that its scheduling team now leaves at 5 rather than 6, with hold times gone. Health + Glow reports more than 1,000 calls a month, a third of them outside office hours, and an allergy and asthma practice about 5,000 a month.

Company figures include 65 percent lower call center operating costs, 300,000 patients helped each month, a tenfold return on AI scheduling in one case study and 93 percent of pilots converting to long term contracts. Call volume describes scale of use. None of the figures says how many calls were resolved correctly, how many needed staff to fix, or how the savings were measured.

Outside AI providers are contractually barred from training on Talkie or customer data, and each must hold a business associate agreement or data processing agreement matching the data it touches. A permission matrix limits which classes of data can reach which AI tool, nothing in the stack may create a voiceprint or biometric template from a caller's voice, and customer data is deleted at termination under documented disposal procedures.

Retention periods exist by data category, but the periods themselves, including how long call recordings and transcripts are kept, are not given. Safety engineering is described in part: material prompt or model changes are tested against datasets before release, including checks that agents stay within clinical scope, and AI quality issues are routed to product review. How a harmful answer to a patient would be caught during live calls is not described.

Business associate status is stated plainly, and the agreement itself is not open to read. Talkie says it acts as a business associate to its covered entity clients, signs a business associate agreement with every US healthcare client, holds agreements with each subprocessor that may process patient information, applies the minimum necessary standard and follows HIPAA breach notification timelines.

Because every US client signs, the agreement does not depend on a plan tier. What a buyer cannot do is read its terms before contracting, including how it binds those subprocessors. The privacy policy on the website is a general notice for site users written to European data protection law, issued by Talkie LLC, and it does not describe how patient information is handled inside the service.

SOC 2 Type II comes with dates and scope: three consecutive examinations by an independent CPA firm, in April 2024, December 2024 and July 2026, covering all five Trust Services Criteria (security, availability, confidentiality, processing integrity and privacy). The security center lists the controls behind them, from anomaly monitoring and encryption to background checks and recurring training.

The reports themselves are released by request to the security team rather than opened from the page. An independent firm runs penetration tests every year on a rolling schedule, with findings tracked to remediation and reviewed by leadership. No HITRUST or ISO 27001 certification is claimed.

After hours, the agent decides which calls are emergencies. When a patient describes an urgent medical situation, Talkie says the AI identifies it as an emergency and transfers the call, pages or texts the on call provider, following the practice's routing. Everything else it handles itself or logs as a case for staff the next morning. Talkie sells that sorting as the point: the agent intercepts callers who press the emergency option for routine needs, passes on only the calls it judges to be real emergencies and shields on call providers from everything else.

Sorting calls by clinical urgency is triage, and no regulatory position or intended purpose statement covers it. The practice sets where emergencies go, not what counts as one; no named protocol stands behind the judgment, and no clinician reviews the calls the agent keeps overnight. Patients are told they are talking to an AI and may ask for a person, though after hours there is no front desk to take them. Daytime work, scheduling, refill requests collected for staff and messages, stays administrative, and no terms of service are posted.

AI governance follows the NIST AI Risk Management Framework, with named controls behind it. A cross functional committee approves every AI tool before adoption, an inventory records each patient facing AI capability with the data it touches, its exposure to patient information and its risk tier, and material changes are evaluated before release, including checks that agents perform consistently across patient groups.

None of those evaluation results are shared. The gap matters for an agent marketed as speaking more than 50 languages, with customers praising its Korean and Spanish service: no accuracy or completion figures are given by language, accent or age group, which is where a voice agent is most likely to misunderstand a caller who then hangs up without complaint.

Every call leaves a record staff can check, and nothing contractual stands behind what the agent tells a patient. Messages and cases are summarized into the EHR, booked appointments appear on the calendar, callers can ask for a person, and AI quality issues are routed to product review.

No error rate for scheduling, refill capture or call routing is offered, and there is no guarantee, service credit or remediation commitment tied to the agent's output. The pilot works as a trial on real calls before a longer contract, which limits a practice's exposure without being a commitment about accuracy. With no terms of service posted, how liability for a misrouted call is allocated is left to each contract.

Two way integration with named record systems is the backbone of the product. Talkie is a certified athenahealth Marketplace partner, listed among the three fastest growing partners for the second quarter of 2026 after athenahealth's own evaluation and testing, and it has integration pages for ModMed, Elation and AdvancedMD. The agent reads schedules and patient data and writes appointments, new patient charts, refill requests and messages back as cases in real time.

Production use shows in customer stories. ESD Pediatric Group runs its agent on athenaOne, creating patient cases and flagging calls for staff follow up, and an allergy and asthma practice on athenahealth routes about 5,000 calls a month through it. Phone system integrations connect the agent to the practice's existing lines, and setup typically takes two to four weeks.

Amazon Web Services hosts the service, and data residency is enforced by market: US client data is processed and stored in US regions, with backups kept in a separate region. Data is encrypted at rest with 256 bit AES and in transit with TLS 1.3, and access to production requires multifactor authentication over encrypted connections.

It is a single hosted service, with no private cloud or on premises option. Tenant isolation between practices is not described, and the model, speech and telephony subprocessors that process calls are mapped per client rather than named, so a buyer cannot confirm before contracting that every processing step stays in US regions. Talkie LLC runs the US service, while the company's own site gives its legal name as Talkie sp. z o.o., a Polish limited company; where the engineers and support staff with production access are based is not stated.

Plans start at $1,000 a month, which covers about 5,000 minutes of agent time, or roughly 2,500 calls resolved, with overage billed at 20 cents a minute. Every practice begins with a paid pilot of about 90 days, tuned with a dedicated implementation team, and then chooses month to month billing or a 12 month agreement at 20 percent off.

The quote is built around call volume, specialty mix and the workflows the agent handles, and it includes implementation, read and write EHR integration, support and ongoing tuning. The pilot fee and how pricing scales above the entry plan are left to the quote, and buying starts with a demo rather than self serve signup. Talkie sets its cost per call, 40 to 80 cents, against $1.64 to $3.29 for a front desk hire.

Independent practices and physician groups are the market, sized as small, medium and large practices, with a dedicated offer for Privia practices. Specialty pages cover dermatology, obstetrics and gynecology, orthopedics, pediatrics, primary care, urology, neurology, ENT, endocrinology, allergy and multispecialty groups, each describing the calls that specialty fields, such as biologic and topical refills in dermatology.

Customer stories back several of those: pediatrics (ESD Pediatric Group, Royal Oak Pediatrics), urology (Urology Group of Southern California), dermatology (Young Skin Dermatology), primary and family care, radiation oncology (MHP Radiation Oncology Institute), vascular surgery and an allergy and asthma practice. Neurology, ENT and endocrinology have pages but no named customer, and hospital and health system call centers sit outside the stated market.