SIAgentNews.com
← All news
DevelopingClaudeAgents

Stretto builds AI voice agent for bankruptcy intake using Amazon Nova 2 Sonic

The company replaced static web forms with a conversational AI agent to collect complex financial data from stressed callers.

· First reported by Amazon Web Services (AWS)

95 percentreduction in per-minute voice cost

The short version

  • Tara uses a speech-to-speech model to handle conversations natively, reducing latency and improving natural flow.
  • The system achieved a 95 percent reduction in per-minute voice costs compared to the company's previous solution.
  • The AI handles complex conditional branching to collect data on assets, debts, and employment history.
  • The solution uses a session continuation pattern to manage interviews that exceed the model's streaming limits.
  • The system includes a dedicated framework for adversarial and compliance testing to handle sensitive financial data.

Why it matters

This demonstrates how speech-to-speech models can handle high-stakes, emotionally charged interactions that were previously too expensive or complex for automation. It shows a shift from simple data collection to sophisticated, rule-based lead qualification in sensitive industries.

Moving beyond web forms

Before implementing AI, Stretto used a standard web form on its bankruptcy site to gather lead information. This method captured basic details like debt amounts and contact info, but often resulted in thin margins for the attorneys who received the leads.

The company faced a dilemma when trying to gather more comprehensive data. While a longer form could capture details like real estate, vehicles, and employment history, most visitors would abandon the form halfway through the process.

Stretto identified that people are generally more comfortable sharing sensitive financial details during a natural conversation. However, hiring and training a human team to conduct these phone interviews was too expensive to be cost-effective for the business.

The architecture of Tara

To solve this, Stretto built Tara, an AI voice agent powered by Amazon Nova 2 Sonic. This is a speech-to-speech foundation model available through Amazon Bedrock that processes audio input and generates speech output in a single step. This design avoids the need to chain separate speech-to-text and text-to-speech components.

By using a single model, the team was able to achieve lower latency and more natural-sounding interactions. This is critical for bankruptcy intake calls where callers may be stressed or hesitant. The architecture also allows the agent to handle interruptions, known as barge-ins, without breaking the flow of the conversation.

The technical setup uses Twilio for telephony and Amazon ECS on AWS Fargate to run the voice bridge. After the call ends, the system uses Anthropic’s Claude 3 Haiku to extract structured data from the transcript and save it to a relational database.

Handling complex data

The scope of the AI's task expanded significantly during the production phase. Tara was designed to handle over 18 intake categories that include conditional branching, meaning the questions she asks change based on the caller's previous answers. This allows for the collection of complex relational data such as multiple debt line items or various vehicle details.

The system also applies attorney-defined rules to sort the leads automatically. Based on the data collected, the system marks a lead as qualified, disqualified, or needing more information. This helps attorneys make faster decisions by focusing on the most relevant cases.

Because the voice agent can gather richer data than a static form, it provides more flexibility for these rules. Qualified leads are then routed to practitioners in the correct jurisdiction, while others are directed to alternative resources.

Overcoming technical limits

One challenge the team faced was the duration of the interviews. A full attorney-designed intake can last between 15 and 20 minutes, which exceeds the current streaming limit for the Amazon Nova 2 Sonic model. To solve this, the team implemented a session continuation pattern.

The voice bridge tracks how long a bidirectional stream has been open. Just before the limit is reached, the bridge opens a new stream using the same system prompt and the conversation history as text. This allows the caller to continue the conversation smoothly without realizing the underlying session was reset.

The team also prioritized emotional awareness during the build. They spent a week testing how the agent responded to different emotional states, such as frustration or hesitation. The goal was to ensure the agent sounded empathetic rather than robotic during these sensitive financial discussions.

Testing for compliance

Deploying an AI for bankruptcy requires rigorous testing beyond basic functionality. Stretto worked with the AWS Generative AI Innovation Center to build a framework covering four categories. These include happy path testing, where cooperative callers complete the full questionnaire accurately.

The team also performed adversarial testing to see how the agent handled contradictory information or callers who refused to answer specific questions. They also conducted compliance testing to ensure the agent correctly delivered legal disclosures and collected necessary consent before handling sensitive data.

Finally, they performed specific barge-in scenario testing. This ensured that if a caller interrupted the agent during a disclosure or mid-question, the AI could recover the context. These tests served as a gate to ensure the system was ready for real-world production use.

Production results

After the pilot, Stretto began optimizing the system based on real caller behavior. They tracked metrics such as call volume, average duration, and profile completeness to identify where callers were dropping off. These insights allowed the team to reorder questions to keep callers engaged longer.

The team also adjusted the conditional branching to prioritize high-value fields earlier in the conversation. They modified the qualification thresholds based on direct feedback from the attorneys who receive the leads. This iterative process ensures the AI remains useful for the human experts who rely on its output.

By moving to Amazon Nova 2 Sonic, the company reported a 95 percent reduction in per-minute voice costs. This shift, combined with the ability to handle complex, multi-phase interviews, allows the company to scale its intake process without a linear increase in human staffing costs.

Sources and links

First reported by Amazon Web Services (AWS). Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news