AI-Driven Voice Automation in Healthcare Telephony : From Voicemail to Voice Agent: AI-Driven Automation of Administrative Telephony Tasks in Swedish Healthcare

Sammanfattning

This thesis investigates whether an AI-driven, audio-native digital voice assistant can support or partially automate selected administrative telephony workflows in Swedish healthcare: (1) prescription renewal and (2) appointment management (cancellation/rescheduling). A mode-based architecture, centered on gpt-realtime, was designed and implemented with mode-specific prompts, scoped tool access, confirmation-before-action policies, and runtime safeguards for turn-taking and silence recovery. The evaluation comprised four phases: automation suitability assessment, spoken information understanding under public switched telephone network (PSTN) telephony conditions using synthetic voice data, task-completion performance evaluation, and user experience (UX) evaluation. Spoken information results showed that two-run OR accuracy (at least one correct run out of two) ranged from 97.9–100.0% for dosage and 74.0–92.0% for medicine names across all evaluated conditions. The task-completion evaluation showed 42/42 passed application programming interface (API) integration executions, and after manual review of flagged cases, all 30 appointment management summaries and all 84 prescription renewal summaries were judged correct. In the UX evaluation, five participants completed all predefined scenarios (20/20 tasks), with category means between 4.30 and 4.87 on a 1–5 scale, while qualitative feedback most often highlighted perceived latency during backend processing. Within the evaluated scope, the results indicate that the system can support selected administrative telephony workflows in a controlled setting within the defined safety boundaries by collecting information, confirming critical details, and producing staff-facing drafts. The next steps for future development include production electronic health record (EHR) integration, broader real-world validation across multiple caller groups and acoustic conditions, implementing measures to protect the system against jailbreaks, and ensuring that the relevant legal, contractual, and data-governance requirements are fulfilled.