Primary interaction model
Keypad choices, fixed menu branches, and short speech inputs mapped to a predefined flow.
Natural-language speech interpreted into intents and structured fields, followed by bounded workflow actions.
Best operating fit
Stable call routing, simple account or appointment lookups, status announcements, and low-variance transactions.
Multi-turn administrative conversations such as confirmation, rescheduling, cancellation capture, waitlist offers, and follow-up qualification.
Caller flexibility
The caller adapts to the menu and selects from the choices the designer anticipated.
The system can interpret more varied wording, but only tested intents and approved actions should be considered in scope.
Workflow change surface
Menu prompts, branch logic, routing tables, integrations, and recordings or text-to-speech content.
Prompts, intent and entity logic, model behavior, tools, knowledge, test sets, language variants, escalation policy, and integrations.
Exception handling
Fallback keys, retry limits, invalid-input branches, queue transfer, callback, or agent handoff.
Confidence and policy thresholds, clarification, tool-error handling, safe refusal, retry limits, and context-rich human handoff.
Integration depth
Often reads or writes a narrow set of fields through deterministic APIs or contact-center connectors.
May orchestrate several reads and writes, which increases usefulness but also expands permission, testing, reconciliation, and rollback requirements.
Language deployment
Each language needs approved prompts, menu parity, routing, recordings or TTS, and usability testing.
Each language also needs speech-recognition, intent, entity, pronunciation, dialogue, escalation, and noisy-call testing; a language label alone is not evidence of production quality.
Audit evidence
Call path, keypress or speech input, routing outcome, transfer, and integration logs.
Transcript or redacted interaction record, interpreted intent, tool calls, confirmations, policy decisions, model and prompt version, handoff, and final disposition.
Failure modes
Menu loops, wrong routing, dead ends, recognition failure, stale prompts, and integration outages.
Misinterpretation, unsupported requests, incorrect tool arguments, model or retrieval failure, latency, inconsistent language behavior, and integration outages.
Cost model
Design and maintenance, telephony, contact-center licensing, recordings or TTS, integration, and agent transfer cost.
Implementation, telephony, speech and model usage, orchestration, monitoring, QA, support, human review, and integration maintenance. Lower cost should be proven with the hospital workflow and volume, not assumed.
Human role
Agents receive transfers and resolve requests outside the menu or system capability.
People own clinical questions, urgent language, identity or consent failures, low-confidence cases, policy exceptions, complaints, and actions outside the approved workflow.