What happened
AWS’s announcement, co-written with Stretto and CI&T, describes Tara, a phone-based AI agent for visitors to Bankruptcy.com. The earlier web form captured debt amount, attorney status, and contact details, but Stretto wanted richer financial information without asking visitors to complete a much longer form. Tara introduces herself, collects consent, and guides callers through a structured bankruptcy interview covering personal information, debts, assets, employment history, and prior filings.
The conversational layer uses Amazon Nova 2 Sonic through Amazon Bedrock. AWS describes Nova 2 Sonic as a speech-to-speech model that processes incoming speech and generates spoken responses in one model rather than chaining speech recognition, a language model, and speech synthesis. The call arrives through Twilio Media Streams over WebSocket, passes through AWS WAF and an Application Load Balancer, and reaches a voice bridge running on Amazon ECS with AWS Fargate. That bridge streams audio between the caller and Nova 2 Sonic. After the call, Amazon SQS and a dead-letter queue support asynchronous processing; Claude 3 Haiku on Bedrock extracts typed fields from the transcript, which are stored in Amazon Aurora Serverless v2. Qualified leads are routed by jurisdiction. The bridge runs continuously across three Availability Zones, and the database is deployed in Multi-AZ configuration.
AWS says the full attorney-designed intake takes 15–20 minutes, longer than one Nova 2 Sonic streaming session can run. Stretto implemented session continuation: shortly before a stream’s limit, the bridge opens a new stream with the same system prompt and tool configuration, passes the conversation so far as text history, switches the caller’s audio after Tara finishes her response, and closes the old stream. The team first deployed the existing 13-field intake as a production baseline, then expanded it over four weeks to more than 18 categories with conditional branching. The expanded schema handles one-to-many records, such as multiple vehicles, properties, and debt items.
The rollout began with a dedicated week of voice-quality checks across calm, stressed, hesitant, and frustrated caller scenarios, measuring latency, interruption handling, and tone adaptation. AWS says that comparison report became the go/no-go gate for integration; the end-to-end integration then took three weeks. After roughly three months in production, Stretto reports an approximately 96% data-extraction accuracy in testing, an average completed-call duration of about eight minutes, a 20% relative increase in conversion versus the web form, and a 95% reduction in per-minute voice cost versus its prior solution. The cost comparison is based on Bedrock on-demand pricing in us-east-1 as of April 2026.
What the archive adds
New Runtime’s earlier coverage described an AWS Samples voice travel concierge for airline customer interactions in Guidance for Travel Concierge using Amazon Bedrock AgentCore and Nova Sonic (2026-10-24). Stretto is a new customer case, not a restatement: it adds a production intake workflow, session continuation for long calls, and reported conversion and cost outcomes.
Trend signal
This supports agent-ready software exposes capabilities: Tara’s voice interface is paired with structured field capture, tool use, post-call extraction, and explicit lead-routing rules. The case also shows that the operational boundary is not just a conversational model; the voice bridge, data schema, and human decision owner shape what the system can safely do. The report is a single vendor-published case, so it is evidence of an implementation pattern rather than proof of a broader industry shift.
Why now
AWS says the team chose Nova 2 Sonic after its previous solution produced time-to-first-audio latency above 500 milliseconds and handled barge-in poorly. The account is timely for teams evaluating speech-to-speech systems because the reported cost reduction and interruption handling are tied to an actual production workflow, not only a model demo.
Why it matters
Sensitive voice workflows fail if a technically capable model cannot keep a caller comfortable, preserve context, or handle correction and interruption. Stretto’s approach separates those concerns: Nova 2 Sonic manages live speech; a continuation pattern bridges session limits; Claude extracts fields after the call; and attorney-defined rules, followed by human review, determine lead qualification. That division makes the system easier to inspect than a single model responsible for conversation, extraction, and consequential decisions.
The rollout sequence is also useful. Test voice quality and interruption behavior before building the integration, establish a limited production baseline, and expand the questionnaire only after observing real calls. Stretto says production data drove changes to question order, branching, and qualification thresholds. Teams can borrow the method while treating the reported conversion and savings as customer-reported outcomes that may depend on their own call mix and economics.
Limits and uncertainty
The conversion increase, extraction accuracy, and cost reduction are reported by Stretto in an AWS customer case; the announcement does not present an independent evaluation or a detailed measurement methodology for those figures. It does not establish how results vary by jurisdiction, caller population, or deployment scale. AWS notes that Bedrock and Nova 2 Sonic are not offered in every Region. The account describes consent, disclosure, security, and adversarial testing controls, but does not establish legal compliance across jurisdictions or quantify error rates for individual extracted fields. The system leaves final lead decisions to attorneys, an important boundary for this use case.
Skill to build now
For a voice workflow with long sessions, prototype the handoff as a first-class test: track stream duration, open the replacement stream before the limit, carry forward the prompt, tools, and text history, then switch audio only after the current response finishes. Test interruptions at disclosures, mid-question, and during answers. Separately, start with a narrow production questionnaire and define the human review point before expanding branching or letting the system route consequential decisions.
Reader takeaway
Stretto’s useful lesson is not simply to swap a web form for a voice model. It is to validate conversational behavior first, roll out a constrained baseline, and keep extraction, qualification, and human authority explicit. Its reported business gains are promising, but teams should reproduce the workflow and measure their own results rather than assume the same economics.
Open question
How will extraction accuracy and conversion change as Stretto expands Tara across more districts and a wider range of callers?