David Plon, founder of Portrait Analytics, argues AI is most useful for triage like idea generation, building context and monitoring a thesis, not for building conviction.
Plon says a raw model should not fill more than about 70% of its context window, because complex tasks degrade, and he tells models to be skeptical of management teams.
Plon says top-down AI mandates often backfire, so firm adoption must happen bottom-up through initiatives that do not change investment process.
Investors Should Use AI for Triage, Not Conviction
AI is most valuable to investors as a triage tool. It surfaces ideas, builds context quickly and watches a thesis, while conviction stays with the human. Results depend on how well the investor briefs the model, how sources match the accuracy needed, and how often new models get tested. Firms adopt it best from the bottom up, because investors must trust it first.
Takeaways
AI is best for triage, not conviction. Plon says AI helps most with idea generation and with building context on a company fast. It also monitors a thesis as news arrives. It does not replace the steps that build conviction, and factual errors are tolerable early in research.
Write prompts like briefs to a smart outsider. Plon writes prompts like a brief to a smart overseas worker without context, giving background, task, desired output, guidelines and domain knowledge. He tells models that management teams are always biased positively, so commentary should be read skeptically.
Match data sources to the accuracy required. For structured quantitative tasks Plon uploads documents, because a model crawling the web will most likely hallucinate. Five 10-Ks run about 500,000 tokens. Gemini's window is near a million tokens and Claude's near 200,000, and he keeps raw models under 70% full.
Spend steady time experimenting with new models. Plon puts about 15% of his time into experimentation and runs a suite of 10 exercises against each new model. Queries cost little, so iteration is real time, and any task a model can do becomes a saved template.
Firm adoption must happen bottom-up. Top-down mandates often backfire, Plon argues, because people must trust AI before relying on it for high-quality decisions. Firm-wide tools like thesis monitoring work because they leave process unchanged. Documenting memos and decisions creates future training data.
Memory is a convenience for now. Plon is near-term bearish on memory, which is limited and expensive in context and far worse than human memory. Reustle suggests investing in the prompt may beat it. Continual learning could eventually make a model that lived through every firm investment far more powerful.
Coding agents thrive where results can be verified. Claude Code and Codex now work for hours on a code base, while Portrait's early GPT-4 agent needed a system prompt of about 30,000 tokens. Plon says code is a verifiable environment, unlike investment research.
Triage examples from the episode
For Expedia, pull data points on hotel demand and online travel agency share.
Review five years of CEO compensation proxies, now a click of a button on Portrait.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
A Bloomberg screen shows guidance beaten every quarter, while full-year guidance was revised down repeatedly.
After the April 2025 tariff announcement, screen for the most exposed companies. The hardest case is encoding a qualitative model of a temporary headwind.