← back
Why Off-the-Shelf AI Doesn't Understand Money — Udi Menkes, Intuit
Takeaway
Financial intelligence requires learning from comparable business decisions and outcomes, beyond supplying a model with one company's records.
Summary
- Intuit examples show frontier models giving risky generic advice despite access to business data, such as expanding a rental portfolio while cash flow is negative.
- Outcome-grounded recommendations account for constraints such as customer concentration and supplier dependencies when choosing between price increases and cost negotiation.
- The speaker reports that over half of advice in a large internal study centered on acquiring customers or increasing revenue, illustrating generic financial reasoning.
- Comparing similar businesses that did and did not take an action is presented as a way to learn from outcomes while addressing the difficulty of causal attribution.
financial-aioutcome-datacausal-inference
Original description
Ask any LLM a financial question about your business. You'll get a fluent, confident, generic answer — one that doesn't truly know your business, or what happened when businesses like yours made that same decision. We build financial AI at Intuit serving 100M+ customers. Our custom LLMs outperform general-purpose models on accuracy while cutting latency in half. But that's the foundation, not the destination. I'll cover where financial intelligence goes when AI stops reporting what happened and starts helping you decide what to do next (and does it for you). Speaker info: https://x.com/menkesu / udimenkes Timestamps: 0:00 - A three year old's theory of money 2:21 - Why off the shelf models fail at money 2:48 - The rental property example 4:34 - The fluent bluff 6:42 - The Princeton million dollar study 9:45 - Context is not experience 11:03 - Correlation versus causation in pricing 13:37 - Building state, action, outcome data 15:17 - Testing it head to head 16:37 - The era of outcome driven finance 17:55 - Three things to remember