Stretto details AI bankruptcy intake, claiming 95% lower per-minute voice costs
An AWS case study describes fully automated intake calls, but the reported savings cover voice costs—not the entire system—and attorneys retain final decisions.
Loading page…
An AWS case study describes fully automated intake calls, but the reported savings cover voice costs—not the entire system—and attorneys retain final decisions.
Listen to this story
Stretto’s November 18 AWS case study documents a production voice agent that turns brief bankruptcy lead collection into adaptive financial interviews, while attorneys retain qualification rules and final decisions. Tara uses Amazon Nova 2 Sonic for calls; a separate Claude 3 Haiku step structures transcript data. Stretto reports eight-minute completed calls, 96% extraction accuracy and 95% lower per-minute voice costs. The figures are company-reported, with no accuracy sample size or scoring method disclosed; the cost claim covers voice costs, not total operating savings.
A full questionnaire can take 15–20 minutes, longer than one Nova 2 Sonic streaming session; the bridge carries conversation history into a replacement session.
Stretto tested voice quality for a week, built the initial production integration over three weeks, then spent four weeks expanding intake to more than 18 categories.
Calls are fully automated and include consent collection; attorney-defined rules sort leads, but practitioners make the final decision.
Stretto’s pitch for automated bankruptcy interviews is richer financial information without the expense of staffing an intake team. In a November 18 AWS case study, the company disclosed its deployed voice agent, Tara, and claimed 95% lower per-minute voice costs than its previous solution. That figure measures voice costs, not total operating savings.
Tara answers calls to Bankruptcy.com and gathers information about callers’ finances, assets and employment. Amazon Nova 2 Sonic, accessed through Amazon Bedrock, powers the conversation. The case study was co-written with representatives from Stretto and implementation partner CI&T; its performance figures are company-reported, not an independent evaluation.
The cost comparison uses Amazon Bedrock on-demand pricing in the US East (Northern Virginia) region as of April 2026. Stretto also reports roughly eight minutes per completed call and about 96% data-extraction accuracy. The case study does not give an evaluation sample size or explain how that accuracy was scored.
Bankruptcy.com previously collected leads through a short web form. Attorneys wanted more detail, including debts, property, vehicles and prior filings. But Stretto found that longer forms drove visitors away, while hiring and training people to conduct phone interviews was too expensive.
The team first tested voice quality for a week, including responses to stressed or hesitant callers and interruptions mid-sentence. It then built the production integration over three weeks, initially collecting the existing form’s 13 fields. Another four weeks went into extending the questionnaire to more than 18 intake categories.
Questions now change with the caller’s answers. A vehicle discussion can capture financing status and monthly payments for each car; employment questions can cover multiple income sources. Production tuning has included moving valuable questions earlier, changing follow-up paths and adjusting qualification thresholds using attorney feedback.
Twilio streams phone audio to a software bridge running on Amazon ECS and AWS Fargate. That bridge exchanges audio with Nova 2 Sonic, which processes incoming speech and produces spoken replies within one model, rather than linking separate transcription, language-generation and speech-generation systems.
After a call, Anthropic’s Claude 3 Haiku extracts structured fields from the transcript, including names, dollar amounts, addresses and vehicle details. Those records go into an Amazon Aurora database. The reported extraction accuracy therefore describes a separate processing step from the voice conversation.
Long interviews required another engineering workaround. The full questionnaire can take 15–20 minutes, exceeding one Nova 2 Sonic streaming session. Before a session ends, the bridge opens a replacement, supplies the conversation history as text and switches the audio after Tara finishes her current response.
The intake calls are fully automated, with no human agents taking calls. Tara introduces herself and collects consent before the interview. Testing included contradictory answers, refusals to answer and attempts to manipulate qualification, alongside checks on disclosures and consent collection.
Stretto says its system applies attorney-defined rules after the call, sorting leads as qualified, disqualified or needing more information. Attorneys make the final decision. Qualified leads go to practitioners by jurisdiction; callers who do not meet the criteria are directed to alternative resources.
Loading discussion...
Join the conversation
Explain where automation feels helpful—or where a human option becomes essential.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.