How Investors are using AI - [Business Breakdowns, EP.240]

Episode Summary

Executive Summary: David Plon explains how AI is reshaping investment research by improving idea generation, thesis monitoring, due diligence, and workflow efficiency without replacing human conviction. He emphasizes prompt craft, experimentation, context management, and documentation as key advantages, and argues that agentic AI and memory will increasingly become powerful research infrastructure for investors.

Main Topics: David Plon’s investing background and why he built Portrait (Priority: 5/5): Plon traces his career from Barclays special situations to Slate Path and Baupost, and says his investing experience revealed persistent research bottlenecks that AI could address. AI use cases in investment research workflows (Priority: 5/5): The conversation focuses on practical applications: monitoring portfolios, triaging new ideas, building context quickly, and surfacing overlooked data points across ecosystems. Idea generation and thesis monitoring (Priority: 5/5): Plon argues AI is especially useful for broadening the search for ideas and monitoring second- and third-order signals around portfolio names, especially for generalists. Prompting, context, and experimentation (Priority: 4/5): He outlines a practical prompting framework, explains when to load documents versus give freedom, and stresses iterative experimentation as models improve. Documentation, memory, and firm-level adoption (Priority: 4/5): Plon says detailed documentation of investment thinking will become more valuable as models gain context and memory, and that adoption works best bottom-up rather than via mandates. Agentic AI and the future of model capabilities (Priority: 4/5): He describes agentic AI as models that reason, reflect, act, and adjust toward goals, and sees software engineering as the leading edge of what will soon matter in research workflows.

Key Arguments: AI is most valuable where research is information-heavy but human time is scarce: idea sourcing, context building, and monitoring. Generalist investors are especially disadvantaged without AI because they cannot manually track all relevant ecosystem signals. AI helps kill weak ideas faster by surfacing existential risks, poor incentives, or misleading management behavior earlier. A major use case is moving deeper-dive work earlier in the process, such as analyzing CEO compensation or historical guidance patterns. Prompt quality improves when the task, rationale, desired output, and domain-specific guidelines are clearly specified. Models should be told to be skeptical of management commentary, since management teams are typically biased positively. Off-the-shelf LLMs are best for exploratory tasks, but structured tasks require documents and/or specialized tools to reduce hallucinations. Investors should spend meaningful time experimenting because model capabilities are evolving quickly and sometimes in unexpectedly jagged ways. Firm-wide AI adoption works best when tools augment existing workflows rather than force everyone into a single process. Documentation of decisions, memos, and research process will become increasingly valuable because it creates machine-readable institutional memory. Memory is useful near term as a convenience layer, but long term it may enable models to behave like a firm-level analyst with accumulated experience. Agentic AI is becoming real because models can now perform longer-running tasks, self-correct, and use tools in iterative loops.

Data Points: Barclays role: Special situations group - Plon’s first investing seat on the trading floor Hedge fund role: Slate Path Capital, generalist - One of Plon’s prior long/short hedge fund roles Hedge fund role: Baupost, generalist on public markets - His most recent pre-Portrait role, covering public equities and distress credit Business school timeline: 2015 to 2017 - Plon says he first seriously considered AI during his Stanford business school years Context window example: ~100,000 tokens per 10-K - Used to illustrate why loading multiple filings can stress model context windows Model context window: ~1,000,000 tokens - Plon cites Gemini as having around this size of context window Model context window: ~400,000 tokens - Plon cites GPT as having around this size of context window Model context window: ~200,000 tokens - Plon cites Claude/Opus as having around this size of context window Experimentation budget: 15% of time - Plon recommends spending roughly this share of time on experimentation with AI tools

Pivotal Quotes: "Imagine you are writing an email to somebody maybe overseas who's going to work overnight and is going to be doing a task for you." — David Plon: Explaining a practical mental model for effective prompt writing "I think spending some 15% of your time on experimentation is really important because... the capability frontier is evolving as well." — David Plon: On why investors should continuously test new model capabilities "The usefulness of them rises exponentially with the amount of context they're given." — David Plon: On why documentation and accumulated firm context matter for AI in investing

Implications: Investors who build disciplined AI workflows will gain speed, broader coverage, and better pattern recognition. The edge will come less from using AI and more from how well firms document, prompt, test, and integrate it into existing research habits.