AI-Powered Voicemail Triage
How a statewide healthcare organization replaced hours of sequential voicemail listening with an AI-powered triage system—deployed into their own Azure tenant for full HIPAA compliance.
100+
Voicemails Per Day
10+
Hours Saved Daily
5
Weeks to User Testing
HIPAA
Compliant by Design
About the Client
A statewide healthcare organization with multiple clinic locations and hundreds of employees, serving patients across Alaska. Their clinical teams—nurses, specialists, and support staff—handle a high volume of patient communication daily, with voicemail as the primary channel for after-hours and overflow contact.
The Challenge
A Hidden Time Sink
Every weekday morning, clinical staff arrived to inboxes full of voicemails—often 30 to 40 on Mondays alone, with over 100 across the organization daily. Each message had to be listened to in full, in real time, with no way to scan or prioritize. Patient messages averaged over a minute each, and many were longer. The only way to review a missed detail was to navigate a clunky button-based interface, rewinding a few seconds at a time.
The process consumed an estimated 10+ hours of clinical staff time per day across the organization—time spent listening, re-listening, manually noting callback numbers, and trying to determine urgency. When a nurse was out sick, covering their voicemail meant getting their PIN. When a patient called multiple departments about the same issue, there was no way to know.
Leadership had no visibility into how voicemails were being processed, no way to measure response times, and no ability to identify when a department was falling behind. The voicemail system was effectively a black box.
Unique Considerations
The Approach
The client's leadership had been exploring how to bring AI into their operations safely—not through off-the-shelf tools that might send patient data to unknown servers, but through infrastructure they owned and controlled. After meeting Delve Group at an AI conference, they identified voicemail processing as the ideal first project: a clearly defined problem with measurable impact that could serve as the foundation for a broader AI strategy.
The fundamental insight was that voicemail is a sequential, opaque medium. You can't scan a voicemail inbox the way you scan an email inbox—you have to listen to each one, start to finish, to know what it contains. AI transcription changes the medium entirely, making voicemail content instantly scannable, searchable, and triageable—presented in a Kanban-style triage board where staff can see all their messages at once and act on the most urgent first.
Before a single real voicemail entered the system, Delve Group built and tested the entire pipeline using AI-generated synthetic data. An OpenAI GPT series model wrote dozens of realistic voicemail scripts—complete with Alaska-area phone numbers, local pharmacy names, and real medication names—and Azure Text-to-Speech converted them into audio files with multiple voice personas. This let the team iterate rapidly in Delve's own Azure environment, demo a working prototype at the very first kick-off meeting, and refine the UI with clinical staff—all before deploying into the client's infrastructure.
A critical architecture decision was deploying the production solution into the client's own Azure tenant rather than hosting it as a SaaS product. This means patient data never leaves their security boundary. Azure OpenAI services fall under Microsoft's Business Associate Agreement, and all storage, processing, and AI inference happen within infrastructure the client owns and controls.
From kick-off to nurses actively testing the system with real voicemails took just five weeks—fast enough that the most frequent feedback from clinical staff during testing was asking when they could start using it for real.
The Solution
Delve Group built a web-based voicemail triage platform that automatically transcribes, corrects, categorizes, and surfaces voicemails in a visual workflow—replacing sequential phone listening with an at-a-glance dashboard.
Intelligent Transcription
Voicemails are automatically transcribed, then refined by a second AI pass that corrects local names, pharmacies, and medical terminology using curated training examples. A coherence score flags garbled messages before AI attempts to repair them.
Visual Triage Board
A Kanban-style workflow organizes voicemails into New, In Progress, and Completed columns. Each card shows the AI-generated summary, category, urgency flag, and caller information—scannable at a glance instead of listened to one by one.
Client-Hosted & HIPAA Compliant
Deployed directly into the client's Azure tenant with private endpoints, Microsoft BAA coverage, and no data leaving their security boundary. The client owns the infrastructure—not the vendor.
Technical Approach
The system uses a three-stage AI pipeline that processes voicemails as they arrive, transforming raw audio into actionable, triaged information:
Transcribe
Voicemails arrive as email attachments from the phone system, are converted to the optimal audio format, and transcribed using Azure Speech-to-Text
Repair & Enrich
An OpenAI GPT series model corrects Alaska-specific names, pharmacies, and medical terms using curated replacement examples—a living knowledge base that improves with every correction
Categorize & Surface
An OpenAI GPT series model classifies each message by type and urgency, extracts caller details, generates a concise summary, and pushes it to the triage board in real time
The Results
Immediate Impact
Hours Reclaimed
Over 10 hours of daily voicemail processing labor replaced with scannable, triaged cards that nurses can act on in seconds instead of minutes per message.
Instant Prioritization
AI-powered urgency flagging and categorization lets clinical staff identify the most important messages immediately—no more listening through 40 voicemails to find the urgent one.
Cross-Department Visibility
When the same patient calls multiple departments, it's now visible in one place—eliminating duplicate callbacks and ensuring nothing falls through the cracks.
Client-Owned Security
Every component—database, storage, AI services—runs within the client's own Azure tenant behind private endpoints. No patient data ever leaves their security boundary.
Operational Benefits
"You're trying to find all the aces in a deck of cards. You can start from the top and flip through one at a time—which is how listening to voicemails works. Or you can lay them all out on the table and just pick up what you need."
Kyle Easterly
Founder, Delve Group
Why Delve Group
HIPAA-First Architecture
Security and compliance designed in from day one—not bolted on. Client-tenant deployment keeps data where it belongs.
Rapid Iteration
AI-generated synthetic data for early iteration, weekly demos, live feedback loops, and in-person testing sessions with clinical staff.
Alaska Expertise
Local knowledge matters when AI models don't recognize your pharmacies, military bases, or neighborhoods.
Strategic Foundation
Building not just a tool but the secure AI infrastructure for a healthcare organization's broader digital transformation.
Project Details
Timeline
Jul - Oct 2025
Industry
Healthcare
Location
Statewide, Alaska
Tech Stack
Azure OpenAI, Azure
Ready to reclaim your team's time?
Let's discuss how AI can transform your organization's communication workflows—securely.