Payment Collections Voicebot

We built a payment collections voicebot on Amazon Nova Sonic. Contactability 10%→22%, recovery uplift 2.5→3.9pp, +$1M/month collected, $650K saved. Scalable, testable, live in production.

Comment

Inspiration

Our inspiration came from a clear operational inefficiency in debt recovery. We observed that human agents were spending thousands of hours on highly structured, repetitive informational calls. This "one-agent-at-a-time" model was not only expensive—costing nearly $700,000 annually—but also suffered from low contact rates and minimal impact. We realized that by leveraging the latest in speech-to-speech AI, we could transform a bottlenecked human process into a scalable, high-performance digital channel.

What it does

The Payment Collections Voicebot is an automated, natural-sounding system driven by Amazon Nova 2 Sonic. It handles outbound collection calls from end-to-end:

  • Automated Outreach: It replaces manual dialing with an intensive, automated system that responds instantly when a customer answers.
  • Natural Conversation: It uses a speech-to-speech model to conduct fluid, human-like interactions in Spanish.
  • A/B Testing: It features a "collections engine" that allows us to test different scripts, tones, and structures to see which recovers the most debt.
  • Real-time Validation: It validates payment dates and commitments on the fly, providing a seamless experience for the customer.

How we built it

A robust, cloud-native architecture focused on observability and scale:

  • Core AI: Powered by Amazon Nova 2 Sonic for low-latency, high-fidelity voice synthesis and understanding.
  • Infrastructure: Deployed on AWS Fargate and managed via Terraform. Our configuration lives in AWS Parameter Store, allowing us to update system behavior and prompts in real-time without redeploying code.
  • Telephony Integration: Integrated with Genesys Cloud using a custom Architect flow and IVR bridge.
  • Monitoring: Implemented an observability stack using Cloudwatch by tracking session metrics, call health, and success rates in real-time.
  • Deployment: Uses Bitbucket Pipelines with optimized caching for rapid CI/CD cycles.

Challenges we ran into

One of our biggest hurdles was infrastructure agility. Initially, using environment variables meant that any small change required a full Docker image rebuild, which was too slow for a production environment. We solved this by implementing a dynamic configuration system via AWS Parameter Store. We also had to tackle the "hallucination" and validation challenge—ensuring the bot correctly understood and verbalized specific payment dates—which we solved by building custom validation and verbalization tools. For example, at the beginning the bot was accepting dates like 30th February.

Accomplishments that we're proud of

Our system has a massive ROI:

  • Cost Savings: Reduced platform costs by $650,000 annually.
  • Operational Impact: More than doubled contactability from 10% to 22%.
  • Financial Recovery: Increased monthly debt collection by $1 Million USD, raising the recovery uplift from 2.5% to 3.9% over the control group.
  • Technical Excellence: Successfully implemented an A/B testing framework that allows for continuous prompt optimization.

What we learned

Our initial production launch revealed several edge cases in prompt engineering and conversational logic. We discovered that users could manipulate the dialogue to "zero out" their debt or provide invalid payment dates, such as February 30th or a date already in the past. These challenges highlighted the critical need for robust server-side validation and dynamic "grounding" to ensure the bot remains tethered to real-world logic.

Furthermore, we identified performance bottlenecks where voice generation latency increased during tool calls (such as while saving a payment commitment). To maintain a seamless client experience, we implemented pre-recorded "filler" audio to bridge these gaps, ensuring the user never experiences awkward silence. Ultimately, we learned that precise prompt engineering and latency management are not just technical details, but fundamental requirements for business success.

What's next for Payment Collections Voicebot

We plan to integrate our Post-Call Analytics project (another entry in this hackathon) alongside our A/B testing framework to perform deep-dive examinations of call recordings. This data-driven approach will allow us to fine-tune our prompts and iterate on conversational structures to achieve superior recovery results over time. Additionally, we will continue to evaluate and adopt emerging models as they are released, ensuring our voicebot remains at the cutting edge—with voices that sound increasingly natural and maintain the lowest possible latency.

Built With

+ 16 more
Share this project:

Updates