About the keynote

Model quality alone does not compress complex expert work; domain intelligence plus well-placed human checkpoints does, and the evidence is in the failure modes, not the demo.

When we set out to build LLM agents for Intuit TurboTax Small Business, the assumption was that better models alone would compress the work. They did not.

What actually cut tax preparation from roughly five days to thirty minutes was pairing long-horizon agents with domain intelligence – structured, machine-readable knowledge of tax rules and small-business context – plus deliberate human-in-the-loop checkpoints at the moments where judgment mattered and hallucination was expensive.

This talk walks through what we measured, what surprised us (token volume was a poor proxy for progress; routing and domain grounding were not), where the agents failed and how we caught it, and what we would stop doing next time.

The result: a system that reached $10M+ in revenue and 5,000+ customers.

In this afternoon keynote, Attendees get an honest account of building AI agents in a regulated, high-stakes domain – evidence, tradeoffs, and failure modes included – rather than a victory lap.

What you will take away

• A concrete example of long-horizon LLM agents working in production: a tax preparation workflow compressed from roughly five days to thirty minutes, reaching $10M+ in revenue and 5,000+ customers in a regulated, high-stakes domain.
• A look inside the build: what Hardik’s team measured, what surprised them — token volume turned out to be a poor proxy for progress, routing and domain grounding did not — where the agents failed, how those failures were caught, and what the team would stop doing next time.
• A clearer view of what actually compresses expert work: not model quality on its own, but domain intelligence — structured, machine-readable knowledge of the rules and the context — paired with human checkpoints placed exactly where judgment mattered and hallucination was expensive.

About Hardik Tiwari

Hardik Tiwari builds AI agents for work where being wrong is expensive. At Intuit he led the long-horizon LLM agents behind TurboTax Small Business, placing human checkpoints at the decisions where judgment mattered and hallucination was costly. He open-sources agentic tooling under PM-OS and writes about domain intelligence — the structured, machine-readable context he argues does more for agent performance than model quality. MBA, Kellogg School of Management.