How Intuit uses LLMs to explain taxes to millions of taxpayers
summary
Intuit uses Anthropic and OpenAI models plus a proprietary Geno platform to generate TurboTax explanations for 44 million returns, combining fine-tuned Claude Haiku (via AWS Bedrock), RAG and graphRAG retrieval, and the in-house tax engine as the ground truth for all numbers. A layered eval pipeline mixes tax experts with automated checks and an LLM-as-judge to catch hallucinations; production challenges include multi‑million dollar model contracts, 3–10s latencies, prompt engineering, and strict safety guard rails.
AI-written summary · about
chapters (7)
- 0:00Introduction to TurboTax
- 1:36Intuit GenOS Platform
- 2:45LLM Solutions and RAG
- 5:03Model Training and Evaluation
- 8:29Automated Evaluation Systems
- 10:08Key Learnings and Latency
- 12:14Question and Answer Session
related talks
29mAccelerating GenAI Innovation: Lessons From Intuit's Agents an...
22mThe Future of Global Taxation in the era of Gen AI
40mInstructLab: Applying Open Source Methods to Building and Trai...
18mGenerative AI in Financial Industry
49mNext-Gen Documentation with AI
39mLeveraging AI and Emerging Technologies for Competitive Advantage
46mBeyond Basic Prompting: Supercharging Open Source LLMs with LMQL's Structured Generation
16mAutomating Technical Privacy Reviews Using LLMs