Artificial intelligence news.
Straight from the original announcements.

Stretto builds Tara, an AI voice agent for bankruptcy intake

November 18, 2026, 19:58By the AI News Bites Editorial TeamSource: AWS
Stretto builds Tara, an AI voice agent for bankruptcy intake
Image: AWS

Stretto has built Tara, an AI voice agent that conducts bankruptcy intake interviews by phone and sorts leads using attorney-defined rules. The system, developed with AWS and CI&T, uses Amazon Nova 2 Sonic through Amazon Bedrock.

Key points
  • Tara conducts bankruptcy intake interviews over the phone.
  • The agent uses Amazon Nova 2 Sonic through Amazon Bedrock.
  • Stretto reports a 95 percent reduction in per-minute voice cost.
  • Claude 3 Haiku extracts structured fields from call transcripts.
  • Leads are classified using attorney-defined rules.

Replacing a short web form with a phone interview

Before Tara, consumers considering bankruptcy on Stretto’s Bankruptcy.com site entered a web form with fields for debt amount, attorney status and contact information. Stretto said these form-generated leads had thin margins because they lacked the detailed financial information bankruptcy attorneys value.

Extending the form to capture debts, assets, vehicles, real estate, employment history and prior filings risked increasing abandonment. A phone interview could collect those details more naturally, but staffing and training a human intake team was cost-prohibitive. Stretto therefore sought to replace the form with an automated voice agent.

When a caller reaches the Bankruptcy.com number, Tara introduces herself and guides the person through a structured, multi-phase interview. It collects personal information, financial details, asset inventories and employment history, while adapting its tone to the caller’s emotional state. After the call, Stretto’s system classifies the lead as qualified, disqualified or needing more information for attorney review.

Amazon Nova 2 Sonic handles the voice conversation

Tara is powered by Amazon Nova 2 Sonic, a speech-to-speech foundation model available through Amazon Bedrock. The model processes speech input and generates speech output in a single model, rather than chaining separate speech-to-text, language model and text-to-speech components. AWS says this architecture is designed to reduce latency and support more natural conversations.

Stretto chose the model over its previous voice solution for cost, latency and interruption handling. According to Stretto, its per-minute voice cost was reduced by 95 percent compared with the previous solution, based on Amazon Bedrock on-demand pricing in us-east-1 as of April 2026. In Stretto’s testing, the earlier solution had time-to-first-audio latency above 500 milliseconds and handled caller interruptions poorly.

The system uses Twilio Media Streams to send raw caller audio over a WebSocket to AWS. A voice bridge running on Amazon Elastic Container Service on AWS Fargate streams the audio with Amazon Nova 2 Sonic. After a call, Amazon Simple Queue Service sends the transcript to an asynchronous processor, where Anthropic’s Claude 3 Haiku extracts structured fields for storage in Amazon Aurora Serverless v2.

Testing voice quality for sensitive conversations

Stretto, AWS’s Generative AI Innovation Center and CI&T took a phased approach, starting with voice validation before building the production integration and expanding the questionnaire.

The team spent a week testing Amazon Nova 2 Sonic against caller scenarios involving calm, stressed, hesitant and frustrated people. The evaluation covered voice quality, time-to-first-audio latency, barge-in handling and whether the agent could adapt its pace and tone when a caller sounded distressed. The resulting comparison report was used as the go or no-go gate for moving forward with integration.

The architecture also uses AWS security and governance controls. Access to Amazon Bedrock models is governed by AWS Identity and Access Management, while traffic can remain on the AWS network through VPC endpoints using AWS PrivateLink. AWS says Amazon Bedrock does not use prompts or model outputs to train models. For Stretto, the combination of automated calls, structured extraction and attorney-defined qualification rules is intended to turn a longer, sensitive intake process into a system that can support attorney decision-making without a human intake team.

Related news

Kimi Code CLI adds native Amazon Bedrock provider, no separate API key needed
Products and toolsNovember 20, 2026, 01:33

Kimi Code CLI adds native Amazon Bedrock provider, no separate API key needed

Moonshot AI has added a native bedrock provider to Kimi Code CLI v2.7.0, letting the open source coding agent reach models on Amazon Bedrock with standard AWS…

AWS brings PolyAI’s Dialog-RSN-1 voice model to SageMaker JumpStart
ModelsNovember 19, 2026, 18:28

AWS brings PolyAI’s Dialog-RSN-1 voice model to SageMaker JumpStart

PolyAI’s Dialog-RSN-1 reasoning model for real-time customer conversations is now available through AWS Marketplace and Amazon SageMaker JumpStart. AWS customers can…

Airbnb Uses Cost per Task to Tune Coding Agents on Amazon Bedrock
For developersNovember 18, 2026, 22:48

Airbnb Uses Cost per Task to Tune Coding Agents on Amazon Bedrock

Airbnb is using cost per task, alongside model evaluations and production monitoring, to select models and manage coding agents running through Amazon Bedrock. The…