# Voice AI Where One Wrong Question Breaks the Rules - Shiv Indap, Upstart & Brooke Hopkins, Coval **URL:** \ **Category:** Content Creators\ **Tags:** humans-in-the-loop, voice, agents, automation, ai-engineer, machine-learning\ **Created:** [31 October 2026 07:00 UTC](https://www.artofsm.art/t/voice-ai-where-one-wrong-question-breaks-the-rules-shiv-indap-upstart-brooke-hopkins-coval/27213 "2026-10-31T07:00:11Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [31 October 2026 07:00 UTC](https://www.artofsm.art/t/voice-ai-where-one-wrong-question-breaks-the-rules-shiv-indap-upstart-brooke-hopkins-coval/27213/1 "2026-10-31T07:00:11Z")
[![](https://www.artofsm.art/uploads/default/original/3X/a/f/af3e0e4113ffbf5786165163c237d2a44491b3c5.jpeg "Voice AI Where One Wrong Question Breaks the Rules - Shiv Indap, Upstart & Brooke Hopkins, Coval") ](https://www.youtube.com/watch?v=3osCbSufv8g) Brooke Hopkins and Shiv Indap discuss the complexities of building and evaluating autonomous voice agents in regulated industries like fintech, highlighting the need for sophisticated evaluation frameworks that combine automated metrics, human review, and continuous improvement to ensure compliance and reliability. Their collaboration between Coval and Upstart demonstrates how structured human-in-the-loop processes and iterative model refinement can effectively monitor, audit, and enhance voice agent performance while adhering to strict regulatory requirements. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [31 October 2026 07:24 UTC](https://www.artofsm.art/t/voice-ai-where-one-wrong-question-breaks-the-rules-shiv-indap-upstart-brooke-hopkins-coval/27213/3 "2026-10-31T07:24:25Z")
In this presentation, Brooke Hopkins, founder of Coval, and Shiv Indap from Upstart discuss the challenges and solutions involved in building and evaluating autonomous voice agents, particularly in regulated industries like fintech. Brooke draws parallels between autonomous vehicles and voice agents, emphasizing the complexity of ensuring reliable, compliant interactions through sophisticated evaluation infrastructure. Coval works with various enterprises to develop evaluation tools that monitor and improve voice agents used in customer support and logistics. Shiv shares Upstart’s experience with automating verification calls in lending, a highly regulated domain. The voice agent asks applicants specific questions to verify their identity and financial information while adhering strictly to compliance rules. For example, certain questions about fund usage cannot be asked if they touch on sensitive topics like medical expenses. The system must also handle nuanced responses carefully to avoid regulatory issues. Shiv highlights the importance of sidestepping questions about loan status to maintain compliance and the challenges of auditing thousands of calls initially managed through cumbersome spreadsheets. To improve evaluation and compliance monitoring, Upstart partnered with Coval to implement a structured human review process supported by automated metrics and LLM-based judges. These judges assess calls for compliance and quality, with human reviewers validating the outputs to ensure accuracy. Initially, agreement between human reviewers and the automated system was low, but through iterative improvements—such as encoding tribal knowledge and refining the judge models—the agreement rate increased significantly, reaching 92%. This continuous feedback loop is crucial for maintaining regulatory compliance and improving agent performance. Brooke elaborates on Coval’s approach to continuous improvement, which involves defining clear metrics aligned with product goals, monitoring live calls for failures, and routing flagged issues to appropriate teams for action. The system supports simulations to test fixes before deploying them in production, reducing risk to users. This end-to-end process—from monitoring and human review to automated fixes and re-simulation—enables enterprises to maintain high-quality, compliant voice agents that can adapt and improve over time. The integration of human judgment ensures that automated evaluations remain accurate and trustworthy. In conclusion, the speakers stress the critical role of robust evaluation frameworks in deploying autonomous voice agents, especially in regulated sectors like fintech. Effective evals not only monitor and score agent performance but also prioritize human review efforts and guide continuous improvement. Changes are carefully controlled through feature flags and compliance oversight, ensuring safe and reliable deployment. This methodology provides a blueprint for launching autonomous systems that meet stringent regulatory requirements while delivering scalable, high-quality customer experiences. ## Useful Links - [Continuous Evaluation and Improvement of Autonomous Voice Agents - Coval Approach](https://www.coval.ai/blog/voice-ai-in-banking-why-evaluation-is-critical-for-compliance-and-customer-experience/) — Explains the continuous improvement loop and evaluation metrics critical to the video content. - [Large Language Model (LLM) as Judge for Compliance Evaluation](https://arxiv.org/html/2608.24314v1) — Key technical method for automated compliance evaluation described in the video. - [Regulatory Compliance in Automated Lending Calls](https://www.360factors.com/blog/automated-regulatory-compliance-management-trends-banking-fintech-services/) — Provides context and explanation of compliance constraints critical to the use case discussed.