This episode is about how AI helps investors work faster without replacing them. David Plon says AI's main value is filtering information and killing bad ideas early, not making final decisions. For example, AI can automatically analyze CEO pay or check if management's past guidance was honest—tasks that used to take hours. He compares using AI to emailing a smart but clueless colleague: give it context and a task, and it works. No specific stocks are recommended; Expedia and Marriott are mentioned as examples of how AI can monitor signals around a company, like hotel industry trends.
At a Glance This edition of Business Breakdowns invites David Plon, founder of Portrait Analytics, to discuss how investors can practically apply AI. The core argument: AI is not a replacement for analysts, but rather an "intelligent filter" and efficiency tool that helps investors screen informatio
David Plon, founder of Portrait Analytics, former buy-side investor at Baupost and Slate Path Capital, now provides AI-driven investment research tools for investment institutions. The core takeaway: AI is not a replacement for analysts, but rather serves as an "intelligent filter" and efficiency lever, helping investors screen information faster, generate ideas, and execute quantitative analysis in the pre-investment process—yet final decisions and deep research still require human judgment.
David Plon believes that AI's greatest incremental value lies not in tracking a single company, but in capturing sparse signals across the ecosystem.
Traditionally, investors track individual stocks' filings, transcripts, and news via watchlists, which is sufficient. However, the real challenge lies in broad-spectrum monitoring of the ecosystem surrounding portfolio companies—among the movements of customers, suppliers, and competitors, only a handful of data points are relevant to your investment thesis. For example, holding Expedia requires knowing what Marriott is saying, but not its full earnings report—you only care about signals related to OTA distribution strategy and consumer travel demand. In the past, this required industry-expert-level information density; now, AI can act as a "smart filter," extracting sparse data points relevant to the thesis from vast amounts of information.
> "There was no smart filter on top of all that data. I think with AI today… you're now able to pick up those data points across that surface area."
David's self-acknowledged uncertainty: The effectiveness of this capability heavily depends on the tool's depth of understanding of the user's thesis. Generic LLMs, lacking context, may still miss key signals.
David Plon emphasizes that AI's core role in the pre-investment phase is to "make bad ideas die faster," while simultaneously advancing analysis typically reserved for deep research into the initial screening stage.
As a former buy-side investor, David notes that upon annual review, the number of ideas that actually enter deep research is extremely low. Many ideas could have been killed at an earlier stage—once insurmountable issues such as existential risks or misaligned management incentives are identified. AI can provide sufficient baseline context, allowing investors to quickly determine whether an idea passes the "initial sniff test," thereby reserving the 12 hours of deep reading for ideas that are truly worthwhile.
More critically, AI enables analysis types originally belonging to deep research to enter the screening pipeline earlier. For example, CEO compensation analysis—previously requiring a review of the past five proxy filings and manually tracking compensation metrics and their weight changes, a labor-intensive task reserved for the deep research phase—can now be completed with a single click on Portrait. This means analysts can "turn over more stones" per unit of time, concentrating creative deep research efforts on ideas that have passed initial screening.
David provides specific, template-able analysis examples:
David Plon distinguishes between two primary applications of AI in idea generation, with the latter being far more challenging than the former.
The first: understanding which companies are exposed to a particular trend or event. For example, following the tariff announcement in April 2025, investors needed to quickly identify which companies were most affected, as well as the second-order effects—namely, which companies had U.S.-centric supply chains while their competitors relied on international supply chains. Modern LLMs contain extensive world knowledge and can directly handle this type of analysis.
The second (more difficult but more valuable): finding complex qualitative ideas that match a "mental model." David cites his own past mental model as an example: searching for companies that were once high-performing with attractive business models but encountered a temporary setback (macro/product cycle/execution misstep), and whose franchise value the market is currently repricing. Such ideas are extremely difficult to find through quantitative screening because short-term numbers all look poor. This requires:
1. Extensive qualitative reasoning
2. Investors themselves often struggle to articulate their own mental models clearly—many times it is an "overall feel" rather than a checklist of attributes A/B/C
Portrait is helping investors answer the question "what does a 10/10 idea look like for you" by analyzing their historical trading records. When the system achieves an accurate match, the effect is striking—"Clear my calendar, I'm going to spend the next week just figuring this out."
David Plon believes that good prompt design is currently the biggest lever for obtaining high-quality output from AI. The core mental model is "imagine you are writing an email to a smart overseas colleague who lacks your background, asking them to work through the night on a task."
Specific structural recommendations:
1. Task + Context: Explain what needs to be done and why. For example, "Build this cost curve because I suspect it is changing, which may signal future pricing shifts"—just like explaining the task background to an analyst.
2. Output Format (optional): Specify the output structure (introductory paragraph → data points → table), but sometimes excessive constraints can limit the model's creative synthesis ability.
3. Task Guidelines: Notes specific to the task. For example, when analyzing guidance, explicitly state "please capture soft guidance that lacks specific numbers"; otherwise, the model defaults to only capturing quantitative guidance.
4. Domain Knowledge Transfer: A set of bullet points encapsulating your analyst experience. For example, "Most of what you will read is management commentary, which always carries a positive bias. Please maintain a skeptical eye"—the model is trained to be "helpful," which often skews it toward positivity. A simple reminder can significantly improve output.
David emphasizes that best practices for prompt design change every three months, as model capabilities and available tools evolve rapidly. The key is not to pursue the perfect prompt, but iteration—start with the simplest prompt, gradually increase complexity, and observe changes in output. LLM responses are instantaneous, allowing real-time feedback and adjustments, which is a huge advantage over human analysts.
> "The cost of sending a single query is trivial. Start simple, start adding complexity as it's helpful. It's really like a two-way dance."
David Plon observes that the most successful AI deployment strategy combines "finding ways that require no one to change their behavior" with "letting individuals build trust through experimentation."
The investment research process is highly personal—forcing everyone to use the same tool or template often backfires. David cites Portrait’s practice as an example:
Key insight: Trust must be established at the individual level. Compared to general-purpose tools (e.g., ChatGPT), software customized for the investment vertical (e.g., Portrait) can accelerate this process by optimizing UX, but ultimate adoption depends on whether individuals feel comfortable and can continue making high-quality investment decisions.
David Plon believes that the importance of documenting thought and decision-making processes will grow exponentially as AI capabilities improve, making this the most undervalued long-term investment.
The utility of a model scales exponentially with the amount of context. As context length expands and agentic reasoning capabilities strengthen, AI will evolve from a "research tool" into an "entity capable of executing research workflows within a company." Its execution ability depends on how much data it has about you and how your company operates.
Specific recommendations:
This data is already valuable for human onboarding, but for future AI, it will become the core IP that makes AI uniquely useful to you. Just as hiring a junior analyst requires significant time to impart your thought process and historical trading context—if this data is captured in real time, future AI can instantly load this context.
> "It's hard to know ex ante which data is going to be used and how, but I think it's a pretty reasonable bet that having that data at a minimum is helpful for the humans, but will certainly be helpful for the machines."
David Plon believes that the application of agentic AI in investment research is at an inflection point, transitioning from "barely usable" to "starting to work," but it remains significantly far from truly replacing analysts.
Definition: Agentic AI refers to models capable of reasoning, reflecting, taking actions, and then reflecting on the outcomes of those actions in pursuit of a goal.
Historical Comparison: In Portrait's early attempts to build agents with GPT-4, the system prompt was as long as 30,000 tokens, requiring extreme constraints on model behavior — because once the model went off course, it lacked the ability to self-correct. Today, models are smart enough to execute longer tasks, use tools, and dynamically update their thinking during the process.
Leading Position in Code: Software engineering (e.g., Claude Code, Codex) is the most successful application area for agentic AI today, because:
Challenges in Investment Research:
David's Assessment: The models already possess the ability to "iteratively reason and arrive at complex answers that require dynamically adjusting plans." The issue lies in the engineering work — how to get the model into the right context so it can execute long-term research like a junior analyst, and eventually like a senior analyst. This is still a buzzword, but there is substantially more substance behind it than a year ago.
> "The models have now become smart enough to do longer running tasks through using tools and updating its thinking process as it goes."
| Position | Guest's Stance | Key Data |
|---|---|---|
| Expedia | Mentioned as a monitoring case | Key investment factors: hotel ecosystem dynamics, changes in OTA distribution strategy |
| Marriott | Mentioned as an ecosystem signal source | The OTA-related comments in its remarks are valuable for Expedia investors |
(Note: This episode is a methodology interview and does not discuss specific position moves or valuations of investment targets.)
1. David Plon: "The biggest value of AI is not helping you find good ideas, but helping you kill bad ideas faster." Most ideas should be eliminated during the initial screening phase. AI provides sufficient baseline context for investors to make judgments before committing to 12 hours of deep research.
2. David Plon: "The mental model for prompt design is—assume you are writing an email to a smart overseas colleague who lacks your background, asking them to complete a task overnight." It includes four layers: task + context, output format (optional), task guidelines, and domain knowledge. The key is not to get it right in one shot, but to iterate.
3. David Plon: "Management guidance credibility analysis—the surface beat rate can be misleading. If full-year guidance is lowered quarter by quarter after being set in Q1, only to be naturally beaten in Q4, this has important implications for forward modeling." This is a pattern recognition task that AI can automate, whereas previously it required manually backtracking 3–4 years of data.
4. David Plon: "The most successful institutional AI deployment strategy is to find a way that does not require anyone to change their behavior." At the firm level, use AI for idea generation, screening, and thesis monitoring (without altering workflows); at the individual level, let people build trust through experimentation and adopt it naturally.
5. David Plon: "Recording the decision-making process is the oil of future AI. Model utility grows exponentially with context volume. In the future, AI will be able to load your firm's entire history of trades and thought processes." The most undervalued long-term investment today—institutions that capture real-time thinking processes (not just polished memos) will have a massive advantage.
6. David Plon: "Agentic AI already works in coding (Claude Code/Codex) and is at an inflection point in investment research—models already have iterative reasoning capabilities, but fragmented context and unverifiable outputs are the main obstacles." Code is a perfect environment (text files, local context, verifiable); investment research is completely different, but the underlying capability is already in place.
7. David Plon: "Model memory is, in the short term, just a shortcut for context loading. In the long term—when models can learn abstract concepts and pattern recognition from experience like humans—it will become the core driver of investment research." Current memory features (e.g., ChatGPT's memory) are far less rich and abstract than human memory.
8. David Plon: "LLM responses are instantaneous, and you can give real-time feedback to adjust—this is a huge advantage over human analysts. You don't have to wait overnight to see results." The cost of iteration is extremely low, and this is the core practical method for prompt design.