AI in Investing with Daloopa's founder Thomas Li

Episode Summary

Executive Summary: The webinar argues that AI is transforming finance, but not by replacing analysts’ judgment. Thomas Lee says AI is strongest at language-based generation, summarization, and comparison across documents, while finance still depends on nuanced, numeric, context-rich work. The biggest gains will come from firms that combine proprietary data, internal notes, and strong workflow design—not from generic chatbots alone.

Main Topics: What AI is actually good at in finance (Priority: 5/5): Thomas frames AI as a prediction/generation system best suited for language, summaries, and cross-document comparison, not for precise numeric modeling or data extraction. The 'context problem' and proprietary data (Priority: 5/5): The main value in finance AI comes from feeding models internal notes, transcripts, filings, consensus data, and other proprietary context so outputs become useful and differentiated. Why firms are building internal AI tools (Priority: 4/5): Large hedge funds, long-only firms, and banks are increasingly building custom tools because foundational models are cheap and interchangeable, but their own data and workflows are not. Human judgment and the future of analyst work (Priority: 4/5): AI will automate mundane tasks and speed research, but analysts still need to handle corner cases, interpret nuance, and make investment judgments that models cannot reliably replace. Crowding, game theory, and short-term stock moves (Priority: 4/5): In public markets, understanding who else is in a trade matters as much as fundamentals. Pod shops increasingly model crowding, positioning, and who the next buyer/seller will be. AI adoption across firm types and seniority levels (Priority: 3/5): Adoption is broad across banks, pods, and long-only firms, with the most aggressive users being the people who have agency and are willing to build iterative internal processes. Where alpha is shifting (Priority: 4/5): Traditional sources of alpha like reading 10-Ks or doing channel checks are less differentiated now. More durable edge comes from risk modeling, context, speed, and combining data sources well.

Key Arguments: AI is fundamentally a language-generation and prediction engine; it excels at summarizing, comparing, and blacklining text-based materials, but not at robust numerical reasoning or building financial models from scratch. The most valuable finance AI workflows are those that compare internal notes against transcripts, filings, and other company disclosures to find inconsistencies or changes in narrative. Generic ChatGPT use is limited in finance because it lacks access to key proprietary sources such as earnings transcripts, SEC filings, investor presentations, and internal research notes. The main economic advantage in finance AI is not model quality alone; it is the ability to inject relevant context and guardrails around the model’s outputs. Large firms are building internal AI capabilities because foundational models are increasingly commoditized, and their own proprietary data creates the real moat. AI will reduce time spent on mundane tasks like transcription, summarization, and model updating, but it will not eliminate the need for human judgment on nuanced investment decisions. More senior investors tend to be more strategic and therefore more willing to adopt AI, while junior staff are often more skeptical because they are used to doing the underlying work by hand. The biggest risk in public-market investing is not just bad analysis but crowding and positioning; successful firms try to neutralize that risk with modeling and hedging. Alternative data and basic channel checks are less differentiated than they once were because many market participants now use them. The strongest edge may come from internal risk models that evaluate analyst skill, forecast accuracy, sector variance, and positioning effects separately.

Data Points: Podcast/webinar timing: Posted last week - Andrew explains the webinar was originally posted the prior week before being shared on the podcast channel. Earnings transcript length: About 15 pages - Thomas uses average earnings calls as an example of text that can be summarized efficiently by AI. Typical manual workflow duration: Two weeks - He says comparing analyst models to company disclosures historically could take a risk associate around two weeks of manual Excel work. Planning/travel research example: One page - AI can condense a large amount of travel research into a one-page plan, used as an example of strong summarization ability. Model generation limit: Words, not numbers - Thomas repeatedly contrasts strong language handling with weak numerical reasoning in current models. Alternative data relevance shift: 2017-2019 - He says alternative data was a major source of alpha during this period but is less differentiated now. Foundational model performance comparison: Four times as good and half as expensive - Thomas says firms can switch models quickly as new ones emerge with dramatically better price/performance. Banking/finance work hours: 100 hours a week - He cites the historical persistence of long hours even after productivity tools like Excel made work more efficient. Public-markets alpha horizon: 3, 5, 7 years - Andrew asks whether there will still be room for humans in public finance over these time horizons. Crowding factor: A modeled risk vector - Thomas says crowded positions are explicitly modeled and often hedged by sophisticated pod shops. Earnings call / conference catalyst window: 7 days - He gives an example of channel checks leading to a conference event one week later as a catalyst. Adoption timeline: 12 months to 36 months - Thomas says meaningful AI integration requires patience and commitment to an iterative process over this horizon.

Pivotal Quotes: "AI is not there to generate, it's not there to extract data, it's there to generate objects." — Thomas Lee: A core framing of what AI can and cannot do well in finance workflows. "The context problem is: how do you build a product that solves the problem for your customers using the most amount of data that you can assemble" — Thomas Lee: He explains why proprietary data and workflow integration matter more than model choice alone. "The most surprising thing is how little AI adoption there is." — Thomas Lee: He argues that despite the hype, actual AI use in many finance workflows is still early.

Implications: Finance AI winners will be firms that combine proprietary data, strong compliance, and thoughtful workflows. Generic AI helps with speed, but durable edge comes from context, risk modeling, and human judgment.