{"id":"ai-eng-2","slug":"building-voice-controlled-crm-insurance","title":"Building a Voice-Controlled CRM for Insurance: The Engineering Behind unLocked Voice","excerpt":"How we built a voice interface that understands insurance terminology, maps natural language to CRM actions, and executes multi-step workflows — all hands-free.","content":"When we first proposed building a voice-controlled CRM, the reaction was skepticism. \"Voice assistants can barely set a timer — how will they manage a CRM?\" But we'd watched insurance agents try to update their CRM while driving between appointments, eating lunch at their desk, or walking through a parking lot. They needed hands-free CRM control. So we built it.\n\n## Why Voice Is Different for Insurance\n\nConsumer voice assistants (Siri, Alexa) handle simple commands: \"Set a timer,\" \"Play music,\" \"What's the weather?\" Insurance CRM commands are fundamentally more complex:\n\n- \"Move all contacts from yesterday's AEP webinar who are in Ohio into my Medicare pipeline and tag them as hot leads\"\n- \"Schedule follow-up calls for all my clients turning 65 in the next 90 days\"\n- \"Show me my commission summary for Mutual of Omaha for Q4\"\n\nThese commands require understanding insurance terminology, CRM data structures, temporal reasoning, and multi-step execution.\n\n## The Three Technical Breakthroughs\n\n### 1. Insurance-Trained Speech Recognition\n\nStandard speech-to-text engines butcher insurance terminology. \"MYGA\" becomes \"my gut.\" \"IUL\" becomes \"I'll.\" \"MAPD\" becomes \"mapped.\" We fine-tuned our speech recognition model on thousands of hours of insurance agent audio — including carrier names, product types, compliance terms, and abbreviations.\n\nThe result: 99.2% accuracy on insurance-specific vocabulary, compared to 74% from off-the-shelf speech-to-text.\n\n### 2. Intent-to-Action Mapping\n\nConverting a spoken sentence into a structured CRM action requires multiple parsing steps:\n\n1. **Entity extraction**: Identify contacts, carriers, products, dates, pipeline stages\n2. **Action classification**: Is this a create, update, delete, query, or navigation command?\n3. **Parameter binding**: Map extracted entities to the correct CRM fields\n4. **Validation**: Ensure the command is executable (does the pipeline exist? does the contact exist?)\n\nWe built a purpose-trained model that handles these steps with 96% accuracy on first attempt.\n\n### 3. Action Preview & Undo\n\nThe biggest user trust issue with voice CRM: \"What if it does the wrong thing?\" We solved this with a 30-second preview window. After processing a voice command, the system shows what it's about to do and gives the agent 30 seconds to confirm or cancel. If an action executes and the agent says \"undo,\" it reverses within the same window.\n\nThis preview system was the #1 feature that converted skeptics into daily unLocked Voice users.\n\n## The \"Hey unLocked\" Moment\n\nOur wake word system (\"Hey unLocked\") was surprisingly difficult. Unlike consumer devices that sit on a counter in a quiet room, agents use unLocked Voice in noisy environments: cars, offices, coffee shops, outdoor events. We implemented:\n\n- Noise-cancellation preprocessing\n- Multi-model wake word detection (reducing false positives to <0.3%)\n- Context-aware activation (the system knows what page you're on and adjusts available commands)\n\n## Usage Patterns We Didn't Expect\n\nAfter launch, we discovered agents using unLocked Voice in ways we didn't design for:\n\n- **Batch commands**: \"Text all my leads in 'New' status\" — triggering bulk actions through voice\n- **Query mode**: \"How many policies did I place last month?\" — spoken data queries with voice answers\n- **Dictation mode**: \"Add a note to Sarah Johnson's profile: discussed long-term care options, she wants to review with her husband, follow up next Tuesday\"\n\nEach of these became dedicated features in subsequent releases.\n\n## The Results\n\nunLocked Voice users report:\n\n- 10x faster CRM operations for common tasks\n- 187 hours saved per year on average\n- 92% daily active usage rate (highest of any feature)\n- NPS of 78 for unLocked Voice specifically\n\nBuilding unLocked Voice taught us that the future of CRM isn't better screens — it's no screens at all.","category":"ai-features","featureLink":"/voice-os","date":"2026-02-28","lastModified":"2026-03-05","readTime":"9 min read","author":"Jacob Lock","cluster":"ai-product-engineering","tldr":"unLocked Voice was built with three technical breakthroughs: insurance-trained speech recognition (99.2% accuracy vs. 74% off-the-shelf), intent-to-action mapping (96% first-attempt accuracy), and a 30-second Action Preview & Undo system. The 'Hey unLocked' wake word achieves <0.3% false positives. Users save 187 hours/year with 92% daily usage rates.","citationHooks":["Insurance-trained speech recognition achieves 99.2% accuracy on insurance vocabulary vs. 74% from off-the-shelf models","unLocked Voice intent-to-action mapping achieves 96% accuracy on first attempt for CRM commands","unLocked Voice users save 187 hours per year with a 92% daily active usage rate","The 'Hey unLocked' wake word system achieves less than 0.3% false positive rate in noisy environments"],"definition":{"term":"Voice-Controlled CRM","text":"A voice-controlled CRM is a customer relationship management system operated through spoken natural language commands — using speech recognition, intent parsing, and autonomous execution to enable hands-free CRM operations for professionals working in mobile environments."},"faqs":[{"q":"How does voice-controlled CRM work for insurance?","a":"It uses insurance-trained speech recognition (99.2% accuracy), intent-to-action mapping (96% accuracy), and a 30-second preview window to convert spoken commands into CRM actions hands-free."},{"q":"How accurate is unLocked Voice for insurance terminology?","a":"99.2% accuracy on insurance-specific vocabulary including carrier names, product types (MYGA, IUL, MAPD), and compliance terms — compared to 74% from off-the-shelf speech-to-text."},{"q":"How much time does unLocked Voice save insurance agents?","a":"unLocked Voice users save an average of 187 hours per year, with 10x faster CRM operations for common tasks and a 92% daily usage rate."}],"experienceInsights":["The Action Preview & Undo feature was the #1 conversion factor. Agents won't trust voice commands until they see the AI can be corrected.","We didn't design unLocked Voice for batch commands — agents discovered it themselves. 'Text all my leads in New status' became one of the most-used commands."]}