Upstart and Coval: Voice AI Evals in Regulated Lending

How Upstart Runs Compliant Voice AI Agents With Coval Evals

Shiv Indap leads voice AI at Upstart, the AI lending company, and Brooke Hopkins is the founder of Coval, which builds evaluation and simulation tools for voice agents. Hopkins previously led the evaluation infrastructure team at Waymo, and Coval raised a $28 million Series A in June 2026. At AI Engineer they walk through one production deployment: a voice agent that handles Upstart's loan verification calls, and the review loop that made it safe to run in a regulated industry.

Why a simple call is hard in lending: About 90% of Upstart applicants are decisioned instantly. The other 10% call in to verify information, and the agent asks three questions: will your income continue for the next three months, how will you use the funds, and do you have outstanding obligations such as alimony or wage garnishments. It first has to confirm the caller is the applicant. Then the rules kick in. If someone says the money is for medical purposes, the agent must not ask the amount. If they say it's for travel to a graduation, it has to ask whose graduation, because that changes the use of funds. And it must not answer questions about loan status or timing. "if things go wrong, it's not just a customer experience issue. It could be a potential regulatory issue, which is worse."

From a Google Sheet to structured human review: At launch, every call was reviewed by a human. The operations team logged reviews in a Google Sheet that grew from five columns to 10, then 15, as people added columns at will, and the data became unusable for improving the agent. Upstart moved to structured review queues in Coval. It also cut its metrics down to the ones that mattered: accuracy, clarity, completeness, policy alignment, correct escalation and CSAT. Indap's advice is "don't be too attached to your metrics."

33% to 92% judge agreement: After each call, an LLM judge scores whether compliance requirements were met, with an explanation, and a human reviews the judge's verdict. At launch, "a human reviewer and a LLM agent or an LLM judge agreed only 33% of the time." Three fixes closed the gap: writing down tribal knowledge that had never been put into the agent, resolving disagreements between the human reviewers themselves, and improving the judge as more cases came in. Agreement is now 92%.

Evals as product definition: "Evals are really the product definition." Hopkins groups metrics into three questions: did the agent reach the goal, did it follow the required steps, and how good was the conversation. Metrics can be LLM judges, heuristics or audio processing. Coval scores every live call to catch known failure patterns, and uses clustering and anomaly detection for unknown ones. Failures go to review queues, so humans look only at the calls that matter, and confirmed failures get routed to the right owner: agent errors to the agent team, compliance violations to compliance, bugs to engineering.

Shift left with simulation, then gate every change: Once a fix is ready, Coval re-runs the failing scenario as a voice-to-voice simulation, calling the agent with different accents, background noise and interruptions, so the fix is proven before it reaches borrowers. Hopkins argues that enterprises go wrong when release gates make it easy to launch once but hard to change a model or prompt later. The end state is a self-healing loop in which agents propose fixes from eval data, but at Upstart "changes are always gated behind a feature flag": engineers sign off on judges, and compliance approves any change that affects compliance behavior.

5 Lessons from Upstart on Voice AI Evals in Regulated Industries

• Narrow scope is a compliance feature - The agent handles verification only and politely sidesteps status questions it isn't allowed to answer
• Validate the judge, not just the agent - LLM judge output is reviewed by humans until agreement is high enough to trust
• Tribal knowledge is the hidden gap - Most of the 33%-to-92% gain came from writing down rules humans carried in their heads, and settling reviewer disagreements
• Review only what matters - Automated scoring routes failed or anomalous calls to humans instead of reviewing thousands of calls a day
• Simulate fixes before production - Voice-to-voice simulation tests changes against accents, noise and interruptions before customers hear them

What Upstart's Approach Means for AI Agents in Regulated Work

Upstart's numbers show that the hard part of a production voice agent isn't the model but the review loop around it: a judge that agrees with experts, rules written down instead of remembered, and failures routed to the team that owns them. For any organization putting AI agents in front of customers in a regulated business, the takeaway is to treat agent evaluation as the product spec and keep humans in the loop on every change, even when fixes are automated.