Stretto AI Voice Agent Uses Amazon Nova 2 Sonic
Stretto replaced web-form bankruptcy intake with Tara, an AI phone interviewer, and reports a 20% relative increase in lead conversion. The Stretto AI voice agent uses Amazon Nova 2 Sonic to collect financial information through conversation. Attorneys retain final decisions, creating a boundary between automated intake and professional judgment.
The November 18, 2026, case study describes work with the AWS Generative AI Innovation Center and technology partner CI&T. Its central comparison is conversational intake versus the web form, rather than an AI system competing with attorneys. Stretto wanted more complete leads without asking visitors to finish increasingly demanding forms.
That makes Tara an enterprise workflow story as much as a voice-model story. The reported gains depend on collecting usable information, preserving corrections, and applying explicit rules afterward. A convincing voice starts the process; reliable records and accountable decisions determine whether the process deserves wider adoption.
The Stretto AI voice agent changes what a lead contains
The underlying business problem starts with the difference between contact information and information an attorney can evaluate. A short form can establish interest, but it leaves substantial intake work unfinished. A longer form requests more detail while placing more work on someone already facing financial stress.
Stretto’s alternative puts that questioning into a phone conversation. Tara gathers information about the caller’s financial circumstances, then passes the resulting record into a separate qualification process. The change concerns both the interface and the depth of information available afterward.
The commercial proposition is a more useful intake record, rather than a voice interface added to an unchanged form.
According to the bankruptcy intake case study, the system applies attorney-defined rules after the conversation. Attorneys make the final decision. That division matters because a conversational answer, a structured financial record, and a professional assessment serve different purposes.
Consider a caller discussing vehicles, employment, and outstanding debts. Those answers need relationships: which payment belongs to which vehicle, or which income source belongs to which employment record. A transcript containing every spoken word does not automatically provide that structure.
Conversation also introduces ambiguity that a form sometimes prevents through fixed fields. A caller can correct an amount, refuse a question, or describe circumstances outside the expected sequence. The system must preserve those distinctions instead of treating every spoken statement as a final answer.
The Tara voice agent therefore faces a demanding version of information collection. It needs enough flexibility to let callers explain themselves and enough discipline to produce records that downstream processes can use. Pleasant conversation alone would leave the original business problem unresolved.
The reported conversion improvement makes this change worth examining, but it does not establish why conversion increased. Voice delivery, question ordering, richer information, and qualification changes can all affect the result. The public account does not isolate their individual contributions.
It also describes roughly three months of production experience, rather than announcing a new model release on November 18. Publication and deployment are separate events. Readers should understand the report as an early operational account of an existing application.
For enterprise buyers, that distinction improves the story’s usefulness. A deployed intake system provides evidence about a specific workflow, even when its results remain company-reported. It offers a narrower and more practical claim than a general declaration that callers prefer AI to forms.
The immediate pressure falls on organizations whose lead funnels stop at basic contact capture. Tara raises a concrete question for those teams: how much valuable information disappears between initial interest and a completed intake?
Web forms now face a different economic comparison
Web forms have a straightforward advantage: people can enter information without maintaining a live conversation. Their limitations become more visible when the required record grows complicated. Every additional question asks the visitor to interpret, organize, and supply another piece of information.
Voice shifts some of that organizational burden into the application. The caller can explain a situation, while the system handles sequencing and follow-up questions. Whether that feels easier depends on the conversation, the caller’s preferences, and the sensitivity of the requested information.
The relevant comparison is the cost of obtaining a usable record, including the work that happens after initial contact.
That measure reaches beyond the expense of running a form or processing a minute of audio. It includes abandoned attempts, incomplete records, correction work, and professional review. A cheaper interaction can still produce an expensive outcome if someone must reconstruct the information afterward.
Stretto reports a 95% reduction in per-minute voice cost compared with its previous voice solution. The comparison uses April 2026 pricing in the US East region. It is a dated comparison between voice implementations, not evidence that Tara’s entire operation costs 95% less.
The distinction also prevents a misleading comparison with the web form. The cost result concerns a previous voice solution; the conversion result concerns the form-based intake experience. Those baselines answer different questions and should remain separate.
A buyer evaluating the approach needs both measures. Lower voice-processing expense can make longer conversations affordable, while better conversion can increase the value of each completed interaction. Neither figure establishes the full return without accounting for implementation, oversight, and downstream processing.
The Stretto AI voice agent also changes the work expected from the receiving attorney. A richer record should offer more context before an initial discussion. However, the recipient still needs confidence that the information reflects what the caller actually said.
That creates pressure on intake vendors as well as internal product teams. An attractive demonstration shows a conversation proceeding correctly. A credible production offering must explain what happens when information conflicts, extraction fails, or the receiving professional questions a field.
This is where the web form remains a serious reference point. Its rigidity can frustrate users, but that same rigidity makes individual fields and required answers easy to inspect. Voice must recover comparable clarity from a less constrained exchange.
For knowledge workers, the broader connection is familiar: captured information becomes valuable only when it remains usable later. A searchable knowledge base addresses a related organizational problem, although sensitive intake introduces different access and review requirements.
The implication is specific rather than universal. Organizations handling complex preliminary interviews now have another interface to evaluate. They still need evidence that conversational collection improves their own completion rates and record quality before changing the entire funnel.
Nova 2 Sonic handles speech while other systems build the record
Amazon Nova 2 Sonic is a speech-to-speech model, meaning it accepts spoken input and generates spoken responses within the same model. That differs from connecting separate transcription, text-generation, and speech-generation components. The architectural choice affects how the conversation moves between listening and responding.
Stretto connects callers through Twilio to a voice bridge, an application that passes audio between the telephone connection and the model. The bridge runs on Amazon ECS with AWS Fargate, services that manage and run application containers. Amazon Bedrock provides managed access to the models.
Twilio’s bidirectional audio streams support receiving call audio and sending generated audio back into the call. A WebSocket, a persistent two-way network connection, carries that exchange. This establishes the live communication path without making the telephony service responsible for the entire intake workflow.
A unified speech model simplifies the conversation path, while the surrounding application still owns the business process.
That separation becomes clearer after the caller hangs up. Claude 3 Haiku, Anthropic’s text model, extracts structured fields from the transcript. Amazon Aurora, a managed relational database, stores the resulting information in records with defined relationships.
The architecture therefore uses different models for different jobs. Nova 2 Sonic supports the live exchange; Claude handles post-call extraction. Attorney-defined rules then support lead classification, with attorneys retaining final decisions.
This division gives evaluators several distinct places to investigate errors. A misunderstood spoken amount differs from a correctly transcribed amount assigned to the wrong field. A correct field evaluated against an unsuitable rule creates another problem entirely.
Treating all those failures as “the AI got it wrong” would hide useful operational detail. The voice, extraction, storage, and qualification stages need separate scrutiny. Improvements in one stage do not guarantee improvements throughout the workflow.
Long conversations introduce another boundary. The complete intake can take 15–20 minutes, while AWS documents an eight-minute connection limit for an individual stream. Its speech session documentation describes renewal and continuation support.
Stretto uses session continuation, a technique that transfers conversation context into a replacement model connection. The caller’s interview can therefore extend beyond the lifetime of an individual stream. Maintaining context across that boundary becomes part of the application’s responsibility.
This matters because remembering an earlier correction is a functional requirement, not merely a conversational convenience. Losing context during a long interview can create repeated questions or inconsistent records. A system that sounds fluent before a connection change still needs testing afterward.
OpenAI offers a relevant comparison through its voice agent architectures, including integrated speech models and chained processing approaches. Those options show that enterprise teams face architectural choices beyond selecting a voice. Stretto’s account does not establish a performance ranking against OpenAI.
The useful competitive question concerns control and workload fit. Teams must decide where they need live responsiveness, inspectable intermediate information, and separate business logic. Tara illustrates one allocation of those responsibilities within an AWS-based application.
The Nova 2 Sonic intake architecture also exposes an important limit to model-centric procurement. Choosing the speech model leaves substantial design work around continuity, record creation, and decision boundaries. Those components determine whether a phone conversation becomes a dependable operational record.
Sensitive conversations make interruption handling a quality requirement
A bankruptcy intake interview asks people to discuss financial hardship. The system must collect information while allowing hesitation, correction, and refusal. That context changes the meaning of voice quality: sounding natural is useful only if the application also handles uncertainty appropriately.
Stretto’s account describes validation for sensitive conversations, including interruptions and callers who change their answers. Barge-in, the ability to interrupt the agent while it speaks, is particularly relevant. A caller needs a practical way to stop an incorrect assumption before the interview moves forward.
The strongest test is whether the system preserves a caller’s meaning when the conversation stops following the expected path.
That requires more than recognizing an interruption. The application needs to distinguish a correction from an additional fact and a refusal from a missing response. Otherwise, an apparently successful conversation can produce an inaccurate record.
Consent and disclosure handling create another boundary. According to the case study, Tara delivers disclosures and collects consent before intake proceeds. These steps belong to the workflow itself, rather than serving as optional introductory language.
An interruption during a disclosure illustrates the difficulty. Resuming the conversation is only part of the task. The system also needs a reliable basis for determining whether the required disclosure and consent sequence was completed.
The public account describes testing and controls, but it does not establish universal legal compliance. A published architecture cannot answer every question about deployment, jurisdiction, data handling, or professional obligations. Buyers need to evaluate their own circumstances rather than treating a customer story as certification.
AWS’s model service card likewise recommends use-case-specific assessment. It places responsibility for surrounding application infrastructure and business logic with customers. That guidance supports a cautious reading of Tara’s results.
Emotional adaptation deserves similar care. A model can vary its delivery in response to speech without possessing a reliable understanding of a caller’s psychological condition. A reassuring tone should not become evidence that the system understood every fact or handled every vulnerable caller appropriately.
The Tara voice agent also needs boundaries around what it does with collected information. The reported design reserves final decisions for attorneys and applies explicit qualification rules separately. Those controls reduce the scope of model discretion without eliminating the need to inspect the information entering those rules.
A wrong input can still influence an otherwise consistent classification process. Human authority at the end matters, but its effectiveness depends on what reviewers can see and correct. The presence of an attorney should not become a substitute for measuring upstream errors.
For teams building similar systems, these are operational questions rather than abstract objections to automation. Can callers correct a record? Can reviewers identify uncertain answers? Can the organization explain why a lead received a particular classification?
The Stretto AI voice agent is most informative when viewed through those boundaries. Its significance comes from assigning different responsibilities to conversation, extraction, rules, and professional judgment. Wider deployment will test whether those boundaries remain dependable under more varied real-world conditions.
The reported gains leave important measurement questions open
Stretto’s headline result is a 20% relative increase in lead conversion compared with its web form. “Relative” matters: the figure describes an increase from the earlier rate, rather than a gain of 20 percentage points. The case study does not provide enough detail to calculate the absolute change.
The company also reports approximately 96% structured-field extraction accuracy in its testing. Completed calls average about eight minutes. These figures describe different aspects of the system and should not be combined into a single claim of overall accuracy.
The results support further evaluation, but they do not establish a controlled benchmark for voice intake.
Extraction accuracy measures the conversion of conversation content into fields. It does not automatically measure the truth of a caller’s statements, the appropriateness of qualification rules, or the quality of an attorney’s eventual assessment. Those outcomes require different evidence.
The distribution of extraction errors also matters. An incorrect secondary detail and an incorrect financially consequential field can have different downstream effects. An aggregate accuracy figure does not show whether errors concentrate in the information reviewers consider most important.
The published account does not provide a detailed field-by-field error breakdown. It also leaves questions about sample size, test composition, and performance across caller groups. Those omissions limit comparison with another intake system, even if the reported average remains useful internally.
Call duration needs equally careful interpretation. The eight-minute figure concerns completed calls, while the full intake can take 15–20 minutes. These are not necessarily conflicting statements, because average completed conversations and the possible length of a full interview describe different things.
However, completed-call duration excludes unsuccessful attempts from that particular measure. Evaluators also need to understand abandonment and partial completion. A short average is encouraging only when it accompanies sufficient information and acceptable completion outcomes.
Conversion presents a related attribution problem. The team describes an evolving workflow, rather than an experiment changing only the speech model. Without a controlled comparison, readers cannot assign the reported improvement entirely to Nova 2 Sonic.
The same caution applies to the 95% per-minute voice-cost reduction. It concerns a previous voice solution and a specified pricing snapshot. It does not establish equivalent savings for telephony, data storage, engineering, review, or the complete cost of delivering a qualified lead.
The account also comes from participants in the deployment, including Stretto and CI&T contributors. Their access makes the architectural description valuable. Their involvement means the commercial results should remain attributed rather than presented as independently audited findings.
None of those limitations makes the story uninformative. Early production evidence can justify a broader pilot and more detailed measurement. It simply supports a narrower conclusion than a general claim that automated voice intake outperforms forms everywhere.
For enterprise buyers, the most useful next question is whether another team can reproduce the gains using comparable definitions. Conversion, accuracy, completion, and cost need explicit measurement boundaries. Without them, similar-looking percentages can describe very different operational outcomes.
Watch expansion, consequential errors, and complete lead economics
The next meaningful evidence should concern deployment breadth, information quality, and total operating economics, in that order. Those signals would test whether Tara’s initial results survive beyond the conditions described in the case study. They would also help buyers separate model capability from workflow design.
First, watch for results from broader deployment across additional districts. The reported conversion improvement comes before broad expansion. Comparable outcomes across a wider deployment would strengthen the argument that conversational intake offers repeatable value.
Expansion alone would establish reach, rather than quality. The stronger signal would combine wider availability with consistent conversion definitions and completion reporting. If outcomes deteriorate, the original result would remain relevant but more clearly limited to its initial setting.
Second, watch for more detailed reporting on consequential information errors. Field-level extraction results would help identify where the system needs correction or review. Information about attorney revisions would add another perspective on how useful the generated record becomes in practice.
A strong result would show that important fields remain dependable as conversations become more varied. A weaker result would show a favorable aggregate hiding recurring errors in critical information. Either finding would be more informative than another undifferentiated accuracy percentage.
Third, watch the complete economics of a usable lead. Per-minute voice savings establish only part of that picture. Buyers also need to understand the effort required to process, review, correct, and route the resulting information.
The strongest evidence would connect operating expense with records that attorneys can actually use. Savings that survive correction work and oversight would support the business case. Savings offset by substantial downstream labor would narrow it.
Tara’s longer-term value depends on whether better conversations produce dependable records at sustainable operating cost.
These are observation points, rather than scheduled announcements or promised disclosures. The public account does not commit Stretto to publishing each measure. Their absence would leave the initial claims less independently assessable, without establishing that the deployment failed.
The Nova 2 Sonic intake story already offers a useful architectural lesson. Voice automation works within a larger system of records, rules, and human responsibility. The model handles an important interaction, while the organization remains accountable for the workflow around it.
That is also why the web-form comparison will remain relevant. Forms offer a clear baseline for information collection, even when users find them burdensome. Voice earns its place when it improves the complete process rather than merely making the opening interaction more engaging.
For teams considering a Stretto AI voice agent approach, start with the evidence behind the claimed improvement. Ask how conversion was defined, which fields required correction, and what happened to incomplete calls. Examine whether the receiving professional obtained a more useful record, rather than assuming conversational fluency supplied one. Then compare the complete workflow with the existing intake process, including review and correction effort. Stretto’s experience offers a concrete reason to investigate voice intake, especially when forms fail to capture necessary detail. Read the architecture alongside the measurement gaps, and use both to define what your own evaluation must establish before broader deployment.



