Stretto, a legal technology provider specializing in bankruptcy services, addressed a critical gap in lead qualification by replacing its web form intake process with Tara, an AI voice agent. Previous web forms captured only basic data and suffered from high abandonment, while manual phone interviews were cost-prohibitive. Tara autonomously conducts detailed phone interviews, extracting structured financial and personal information in real time, and classifies leads according to attorney-defined criteria.
The core technology powering Tara is Amazon Nova 2 Sonic, a speech-to-speech foundation model accessible via Amazon Bedrock. Unlike conventional pipelines that chain speech-to-text, language understanding, and text-to-speech modules, Nova 2 Sonic integrates these steps into a single model. This design reduces latency, enables more natural conversation flow, and supports interruption handling necessary for callers who interject or correct themselves mid-dialog.
The system architecture routes telephone calls through Twilio Media Streams to an AWS-hosted voice bridge leveraging Amazon ECS on Fargate. Audio streams interface directly with Nova 2 Sonic for real-time bidirectional conversation. Post-call transcripts are processed by Anthropic's Claude 3 Haiku model on Amazon Bedrock to extract typed structured data, which is stored in Amazon Aurora Serverless. Security controls include AWS PrivateLink VPC endpoints and Identity and Access Management to ensure data privacy and compliance.
Stretto emphasized validating voice quality for empathetic handling of emotionally sensitive bankruptcy discussions. Evaluations tested voice latency, natural tone adaptation to caller emotional states, and robust interruption (barge-in) handling, which Nova 2 Sonic notably outperformed previous solutions. System design includes session continuation techniques to maintain conversation context across the model’s streaming duration limits, enabling comprehensive interviews lasting 15 to 20 minutes.
Benefits realized include a 95% reduction in per-minute voice costs compared with the prior voice solution, about 96% accuracy in structured data extraction, and an average call duration of eight minutes. Importantly, Stretto observed a 20% relative increase in lead conversion rates compared to the web form baseline, indicating stronger engagement and richer data capture driving more qualified leads.
Responsible AI practices underpin Tara’s deployment. Consent and disclosure protocols precede data collection, and structured rule-based lead qualification ensures no autonomous consequential decisions are made by the model. Adversarial and compliance testing were performed to guarantee system resilience against contradictory or incomplete caller input, preserving fairness and data integrity.
While deployed in a phased approach starting with a limited questionnaire, the system later expanded to cover over 18 intake categories with complex relational data. The architecture’s modularity and AWS managed services facilitate scaling and secure operation. However, Amazon Bedrock and Nova 2 Sonic’s availability is region-dependent, and the model streaming session duration limits require careful design workarounds.
Stretto’s project exemplifies how integrated speech-to-speech AI can effectively handle sensitive financial interviews at scale, improving lead quality and reducing operational costs. The success demonstrates the viability of voice AI agents for sensitive use cases when combined with robust cloud infrastructure, privacy safeguards, and iterative validation-based development.
Source & further reading
How Stretto built an AI voice agent with Amazon Nova 2 Sonic ↗Source published 2026-11-18. Claims and limitations should be checked against the original announcement or research.