Stretto Replaces a Web Form With a Voice Agent: Cost per Voice Minute Drops 95%
Stretto collected bankruptcy leads with a short web form, but attorneys need details on debts, vehicles and real estate. People abandon long forms, and hiring live operators is expensive, so calls on Bankruptcy.com are now answered by a voice agent named Tara. A voice minute costs 95% less than with the previous solution. Takeaway: if a form suffers from drop-offs, price a voice interview by the minute.
AI-processed from AWS Machine Learning Blog; edited by Hamidun News
Picture a website form where someone considering bankruptcy enters the amount of their debt, ticks whether they have an attorney, and leaves a phone number. That is how Stretto's Bankruptcy.com worked.
For attorneys, such a lead is nearly useless: they need debts, assets, vehicles, real estate, work history and prior bankruptcies. Stretch the form out, and people abandon it halfway. Hiring and training operators to ask all of this over the phone is too expensive.
So Stretto, together with the AWS Generative AI Innovation Center and the integrator CI&T, built Tara, a voice agent running on Amazon Nova 2 Sonic. By the company's estimate, a voice minute costs 95% less than with the previous solution (at Amazon Bedrock on-demand rates in us-east-1 as of April 2026). For developers, it is an example of how to replace a long questionnaire with a live conversation without being ruined by per-minute costs.
How Tara Handles a Call
A person dials Bankruptcy.com, Tara picks up, introduces herself, obtains consent to the conversation and starts the interview. She collects personal details, finances, a list of assets and work history, adjusting her tone to the caller's mood. Twilio takes the calls: it opens a WebSocket and streams raw audio to AWS. The audio passes through the AWS WAF firewall and a load balancer to the "voice bridge." The bridge consists of containers on Amazon ECS Fargate that run continuously across three availability zones. It relays audio in both directions to Amazon Nova 2 Sonic through Amazon Bedrock.
Nova 2 Sonic is a speech-to-speech model. It listens and replies in voice on its own, with no chain of speech recognition, a text model and speech synthesis. By design, this lowers latency and makes the conversation feel more natural. With Stretto's previous solution, by its own measurements, the first audio arrived after more than 500 ms, and the pauses felt awkward. It also handled interruptions poorly, and people under stress often interrupt and correct themselves.
After the call, the transcript goes into an Amazon SQS queue with a dead-letter queue. Then Claude 3 Haiku in Bedrock extracts fields from the text: names, amounts, addresses, vehicle details. These are stored in Amazon Aurora Serverless v2. The model does not decide what happens to a lead. Rules set by attorneys flag it as eligible, ineligible or needing clarification. Eligible leads go to attorneys by jurisdiction, and other callers are pointed to other resources. The final word stays with the attorney.
What Is Worth Borrowing
First, test the voice before writing code. The team spent a full week testing Nova 2 Sonic on scenarios with calm, tense, hesitant and irritated callers. They looked at latency, interruptions and shifts in tone. The final report became the condition for moving on to integration.
Second, work around the session limit. A full interview runs 15–20 minutes, while a single Nova 2 Sonic stream lives for less. The bridge tracks stream time and, shortly before the limit, opens a new one with the same system prompt and set of tools. The conversation history is passed to it as text. When Tara finishes her sentence, the audio switches to the new stream and the old one closes. The caller does not notice a break.
Third, roll out in stages. First, the team built the integration in three weeks, and the agent collected the same 13 fields as the web form. That gave them a baseline. Then, over four weeks, they expanded the questionnaire to 18+ categories with branching: questions depend on earlier answers. The database grew into a connected model in which one person can have several vehicles, properties, debts and income sources.
Numbers and Caveats
Stretto reports a 20% increase in conversion compared with the web form over three months, field extraction accuracy of about 96%, and an average completed-call length of about 8 minutes. Treat these figures with caution. The article was published on the AWS blog together with a customer and partner, and there is no independent verification. The 95% savings figure is calculated at rates for a specific point in time and region. There is no comparison with other voice models; all that is known is that Nova 2 Sonic outperforms Stretto's previous solution. In addition, Amazon Nova 2 Sonic and Amazon Bedrock are not available in every AWS region.
Practical takeaway: if your form runs into people refusing to fill in long fields, a voice interview may yield more complete data. It pays off when the per-minute price is low and when decisions are moved out of the model into explicit rules. Build the session length limit into the architecture from the start rather than discovering it mid-conversation.
- Nov 19, 2026Anthropic goes public and raises $81.3 billion, beating SpaceX's $75 billion record
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
- Nov 19, 2026Google's Gemini 4 Neon runs twice as fast as Carbon and costs 17% less per task
- Nov 19, 2026The same Gemma 4 wrote a backend that passed 80% of checks in opencode, but only 41% in OpenHands
Key Questions Answered
How Tara Handles a Call?
A person dials Bankruptcy.com, Tara picks up, introduces herself, obtains consent to the conversation and starts the interview. She collects personal details, finances, a list of assets and work history, adjusting her tone to the caller's mood. Twilio takes the calls: it opens a WebSocket and streams raw audio to AWS. The audio passes through the AWS WAF firewall and a load balancer to the "voice bridge." The bridge consists of containers on Amazon ECS Fargate that run continuously across three availability zones.
What Is Worth Borrowing?
First, test the voice before writing code. The team spent a full week testing Nova 2 Sonic on scenarios with calm, tense, hesitant and irritated callers. They looked at latency, interruptions and shifts in tone. The final report became the condition for moving on to integration. Second, work around the session limit. A full interview runs 15–20 minutes, while a single Nova 2 Sonic stream lives for less.
What are the key details on Numbers and Caveats?
Stretto reports a 20% increase in conversion compared with the web form over three months, field extraction accuracy of about 96%, and an average completed-call length of about 8 minutes. Treat these figures with caution. The article was published on the AWS blog together with a customer and partner, and there is no independent verification. The 95% savings figure is calculated at rates for a specific point in time and region.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.