Retanr: AI Voice Agents for Legal Intake
Real-time AI voice agents for employment law intake. Structured data extraction with compliance guardrails and human-in-the-loop escalation.
The Problem
Employment law firms lose potential clients in the intake process. Calls go to voicemail. Paralegals are overloaded. By the time someone calls back, the lead has moved on. Firms need 24/7 intake without 24/7 staffing.
What I Built
AI voice agents that handle initial employment law intake calls. The system asks the right questions, extracts structured case data, and routes qualified leads to attorneys—in real time.
- Real-time voice interaction — Natural conversation, not a phone tree. Callers describe their situation; the agent asks follow-up questions based on employment law categories.
- Structured data extraction — Every call produces structured output—case type, employer, timeline, key facts, urgency level. No one has to listen to recordings and fill out forms.
- Workflow integration — Extracted data flows directly into the firm’s case management system. Qualified leads get flagged for immediate attorney callback.
What Accuracy Means Here
This isn’t a chatbot—it’s a system that needs to correctly categorize legal situations. I measure:
- Entity extraction accuracy (names, dates, employers, case types)
- Classification accuracy (wrongful termination vs. discrimination vs. wage theft)
- False negative rate (leads that should have been flagged but weren’t)
Human review catches the rest. The goal is “better than a voicemail box,” not “replace a lawyer.”
Real-Time Constraints
Voice AI has a hard constraint: latency. If the system pauses for 3 seconds, callers hang up.
- Speech-to-text — Deepgram streaming (sub-200ms)
- Processing — Structured extraction runs on partial transcripts as the conversation happens
- Response generation — Pre-computed response patterns for common scenarios; LLM for edge cases
- End-to-end target — Under 2 seconds from caller speech to agent response
Human-in-the-Loop
- Sensitive topics (e.g., harassment, retaliation threats) trigger immediate escalation to a human
- All AI-extracted data is flagged for attorney review before any action is taken
- Callers can request a human at any point
- System confidence scores determine escalation thresholds
Integrations
- Deepgram for real-time speech-to-text
- Claude for structured extraction and response generation
- Webhook integration with case management systems (Clio, PracticePanther)
- SMS follow-up for callers who need to provide additional information
- Audit logging for compliance review
Compliance Considerations
- Call recording consent — System announces recording at call start (state-specific)
- Data handling — Extracted data encrypted at rest and in transit
- Retention policies — Configurable per-firm based on jurisdiction requirements
- Attorney-client privilege — System does not provide legal advice—it captures intake data
Building voice AI or agent systems? Let’s talk.