Syllable

The models drive the conversations, and much of what Syllable sells is the platform around them. The company now presents itself as a neutral platform to build, run and optimize AI agents across industries, with healthcare agents (a healthcare receptionist that schedules, takes refill requests and answers office questions; an outbound medical records agent; call routing and appointment scheduling) among its showcases.

Conversations run on large language, speech recognition and speech synthesis models from OpenAI, Anthropic, Google, ElevenLabs and Deepgram that the customer selects, while Syllable supplies the agent builder, orchestration, telephony, multi-cloud runtime, gateways to those model providers, and monitoring. Without the models there are no agents, but the models are bought in and swappable; the platform is what the company builds.

The oversight structure is published and one part is missing. The healthcare receptionist agent schedules, collects refill details and routes them to the clinical workflow, and answers office questions; when a request is outside its scope it transfers the caller to a human operator with the patient's identity, request and collected details, so the patient does not repeat themselves.

Escalation routing is set on the platform by time of day: in business hours to the front desk, after hours urgent clinical questions to the on-call provider while scheduling and refill requests wait for the next business day. The platform offers human-in-the-loop controls, approvals and versioned configuration, and the customer owns the escalation policy. What is missing is how the agent recognizes that a clinical question is urgent, and what it does with a clinical question it judges routine after hours.

The model suppliers are named down to the version, and the customer chooses among them. The pricing page lists the language models an agent can run on (OpenAI GPT-5.2, GPT-5.1, GPT-5, GPT-4.1 and smaller variants; Anthropic Claude Opus 4.6, Sonnet 4.6 and Haiku 4.5; Google Gemini 1.5 Pro and 2.5 Flash), speech synthesis (OpenAI, ElevenLabs, Google Wavenet and Chirp3-HD) and speech recognition (Deepgram Nova 2 and 3, Google STT v2).

The platform describes versioned configuration, side-by-side comparison of prompts, models and tools, labeled test interactions, and published release notes. What stays unstated is which model a given healthcare deployment runs, and no training data characterization applies, since the models are third-party.

The model layer is named, supplier by supplier. The pricing page lists the third-party models an agent can use and passes their costs through at zero markup: language models from OpenAI, Anthropic and Google, speech synthesis from OpenAI, ElevenLabs and Google, and speech recognition from Deepgram and Google, each with its version and rate; the platform runs on AWS, Google Cloud, Azure and Oracle Cloud, and Twilio is named for SMS and SIP.

What is missing is the binding part: no subprocessor list states which of these a healthcare deployment actually uses, and no retention or training terms for patient conversations at each provider are published.

A named deployment with dated outcome figures published by the customer itself, and newer vendor figures. Houston Methodist's own case study of its COVID-19 vaccine hotline reports 91 percent of calls resolved through the voice assistant, an average of 7,500 vaccine calls a day with a peak of 14,583 in a day and 3,500 in an hour, and callers connected to a live agent or nurse when appropriate.

Syllable's customer stories report Houston Methodist scheduling agents cutting average handle time by more than 50 percent and Virtua Health deploying scheduling agents in three weeks, and the home page reports more than 5 million calls automated a year, 10.9 times return on outreach and 100 percent of calls answered by AI within eight weeks, without naming the customer, period or method. The strongest figures come from a 2021 single-purpose hotline rather than the current platform.

General assurances are published; the questions AI raises are not answered. The home page describes encryption, continuous monitoring, strict access controls, regular penetration testing and privacy by design. Conversations pass through third-party language, speech recognition and speech synthesis models the customer selects (OpenAI, Anthropic, Google, ElevenLabs, Deepgram), billed at cost, and the platform records calls and produces transcriptions as metered services. Nothing the vendor publishes states how long recordings and transcripts are kept, whether those providers retain or train on patient conversations, or de-identification.

No HIPAA status is stated on the pages that sell the healthcare agents, and nothing establishes the contract that would cover them. The home page lists SOC 2 Type II, HITRUST e1, GDPR and the EU AI Act; the healthcare receptionist page describes identity verification and EHR connection but says nothing about business associate status. The trust center, linked from the site, and the privacy policy in the footer may say more.

Self serve accounts start on a pay as you go card payment, and there is no word on which tier, if any, carries a business associate agreement. Every conversation passes through third-party speech and language providers, each of which would need to be covered.

SOC 2 Type II and HITRUST e1 are named as the company's own, with the artifacts behind a trust center. The home page lists SOC 2 Type II (linked to the AICPA's SOC page), HITRUST e1, GDPR and the EU AI Act, and describes encryption, continuous monitoring, strict access controls and regular penetration testing; a Trust Center at trust.syllable.ai is linked from the navigation. HITRUST e1 is the entry-level one-year assessment, below i1 and r2. The SOC 2 period and scope, the penetration testing cadence and whether reports can be opened without a request are not stated on the public pages.

The healthcare agent sorts patients' clinical questions by urgency after hours, and no regulatory position is stated anywhere the company publishes. The healthcare receptionist page explains that after hours 'urgent clinical questions go to the on-call provider while scheduling and refill requests queue for the next business day', set through platform routing rules, and that within business hours out-of-scope requests transfer to the front desk.

A clinician takes the calls the agent treats as urgent; the agent alone decides which clinical questions are not, and no named triage protocol is described behind that decision. The same page suggests customers build post-procedure follow-up agents that flag concerning symptoms for clinical review. Scheduling, refill collection and office information are administrative.

Evaluation tooling is offered and alignment with the EU AI Act is claimed; no evaluation of the healthcare agents is published. The platform describes distributed tracing, structured logging, metrics including speech accuracy and conversation success, labeled test interactions and review workflows, side-by-side comparison of prompts and models, versioned configuration, human-in-the-loop controls and auditability, and lists the EU AI Act among its compliance frameworks.

Those let a customer govern its own agents; no results, per-group performance or bias assessment for the healthcare agents are published. The platform supports multiple languages through language groups, and nothing reports how recognition or routing performs by language, accent or age.

Mechanisms exist to trace and review what an agent did, and nothing stands behind its output. The platform records distributed traces, structured logs and conversation analytics, supports labeled test interactions and review workflows, keeps configurations versioned, and passes full conversation context to a human on transfer, so a wrong booking or a misrouted clinical question can be reconstructed.

The published 99.99 percent platform uptime answers for availability, not results, and no error, misroute or intent-accuracy rate is published for the healthcare agents. No warranty, remedy or service credit attached to an agent's output is described.

One record system is named, with scheduling written into it. The integrations hub lists Epic EHR alongside Twilio SMS and SIP trunking, Salesforce and Google Drive, and the healthcare receptionist page says production agents connect to scheduling systems to check availability and confirm appointments in real time and use Epic for patient verification. Houston Methodist's own case study describes self-service scheduling through the voice assistant.

The platform connects to a customer's existing telephony over SIP rather than replacing it, and is listed on the Microsoft Marketplace. No other record system is named on the vendor's pages, and no Epic marketplace listing is cited.

Hosted across several clouds, with location not committed. The platform runs on multi-cloud infrastructure across AWS, Google Cloud, Azure and Oracle Cloud with policy-based routing, failover and elastic scaling, and connects to the customer's telephony over SIP. The model providers a customer selects process conversation content on their own infrastructure. No region, residency commitment or tenant isolation model is published, and 'run anywhere' with automatic failover means processing can move between clouds and providers.

Published rates, a stated unit of charge and a route to start without a sales conversation. The pricing page offers a free trial (limited minutes, three-minute call cap, no card) and metered pay-as-you-go starting with a $150 prepayment and no annual commitment.

Platform rates are published per unit: enterprise voice at $0.06 a minute, SIP connection, recording and transfer per minute, text per million characters, tools per million tokens, transcriptions per minute and workflow invocations per call; third-party language, speech synthesis and speech recognition models are passed through at cost with zero markup, with each model's rate listed.

Standard support is included; enterprise support carries an annual fee waived above $500,000 in annual commitments, and enterprise pricing is available on request. What the rates do not settle is the total for a health system's call volume once model and telephony costs are added, and which tier carries a business associate agreement.

Built for front desk and access work: a healthcare receptionist agent that schedules, collects prescription refill details and routes them to the clinical workflow, and answers office questions across voice, SMS and chat; an outbound medical records agent; appointment scheduling and call routing use cases; and appointment reminders.

The company now sells a general agent platform across industries (home services, insurance, mortgage, retail and customer support showcases sit beside healthcare), so healthcare is one market rather than the whole business. Healthcare validation sits behind part of it: Houston Methodist (its own published case study) and Virtua Health are named, and the company has reported New York-Presbyterian as a customer. No specialty coverage is described, and coverage is configured per customer.