How investors are using AI in their research process

Executive overview

Most investors have more information available than they can process — the constraint is triage, not access. AI addresses this by expanding the surface area a generalist can cover without replacing the conviction-building work that requires human judgment.

David Plon (founder of Portrait Analytics, ex-hedge fund investor) walks through three layers of the research process — position monitoring, pre-buy diligence, and idea generation — and the specific AI use cases that deliver the most lift at each stage.

The core insight: AI's highest value is surfacing patterns that were previously too painstaking to extract — not replacing the reasoning that builds conviction.

Position monitoring and ecosystem signals

• Staying on top of individual names is largely solved; the real gain is casting a wider net across adjacent ecosystems
• Example: owning Expedia means needing signals from Marriott, Hilton, and other OTAs — not their full filings, just the relevant data points
• AI can now apply a smart filter across that broader surface area, giving generalists context previously available only to sector specialists
• Historically, being late to recognise a weakening end market was a common failure mode for generalists — this closes that gap

Pre-buy research and idea triage

• AI helps kill ideas faster by surfacing existential risks or misaligned incentives early, before deep research investment
• CEO compensation analysis (mapping metrics and weighting across five proxies) is now a one-click task rather than a manual deep dive
• Management guidance credibility — tracking soft and hard guidance across multiple years and comparing to outcomes — is now easy to systematise
• Pattern: management that kitchen-sinks in one quarter beats guidance; management that revises down every quarter signals a modelling adjustment
• Two categories of pre-buy checks: templatisable screens (compensation, accounting flags) and qualitative pattern recognition (turnaround setups, franchise value dislocations)

Idea generation and mental model matching

• AI is useful for identifying companies exposed to a specific trend or second-order tariff effect
• Harder problem: finding ideas that match a nuanced, qualitative mental model (e.g. temporarily impaired high-quality franchises)
• The challenge is partly technical and partly linguistic — many investors struggle to articulate their own mental models precisely enough to query
• When it works, the output is a fully formed pitch that immediately clears the calendar

Prompt writing that works

• Frame prompts as an email to a smart, context-poor analyst working overnight
• Structure: background context → task and why → desired output format (optional) → specific guidelines → domain knowledge
• Domain knowledge example: explicitly instruct the model that management teams are positively biased and to apply scepticism
• Constraining the output format can help for structured tasks; leave it open for exploratory work
• Iterate in real time — the cost of a query is trivial, so start simple and add complexity as needed

Context windows and document loading

• Context windows have plateaued (Gemini ~1M, GPT ~400K, Claude ~200K) but models use context more intelligently now
• Needle-in-a-haystack retrieval (find a specific figure) is reliable; tasks requiring simultaneous attention across a long context (build a three-statement model) are more fragile
• When using a dedicated tool that pre-loads documents, don't hold back — provide the full corpus and let the model choose what to use
• With raw off-the-shelf models, avoid using more than ~70% of the context window for complex tasks; break the task down instead

Experimentation as a durable skill

• Dedicate roughly 15% of time to experimentation, even when a "good enough" method already exists
• Maintain a personal suite of 10 benchmark tasks; run new models against them to calibrate the capability frontier
• The frontier is jagged — capabilities appear unevenly, and early discovery compounds
• Experiment by progressively layering complexity into prompts and observing how sensitivity to ordering, specificity, and context changes

Fund-level adoption

• Top-down mandates to change research workflows tend to backfire; conviction-building processes are personal
• Most successful adoption: firm-level initiatives that add value without changing individual workflows — bespoke idea screens, thesis monitoring feeds
• Thesis monitoring in particular requires no behaviour change — new relevant data points appear; ignore them if not relevant, act if they are
• Trust in AI tools has to be built bottom-up, one researcher at a time

Documentation as long-term infrastructure

• Models become exponentially more useful as context increases; a firm's documented thinking is high-value training data
• Capturing not just formal memos but informal reasoning — quick blurbs after earnings, notes from management meetings — creates a richer signal than polished write-ups alone
• As agentic AI matures, a model that has "lived" every investment a firm has made will be far more powerful than one starting from scratch
• Hard to know ex ante which data will matter; the safe bet is to capture everything

Agentic AI and what's next

• Agentic AI: the model reasons, takes action, reflects on results, and adjusts — iteratively pursuing a goal
• Code is the leading edge: constrained environment, local context, and binary verifiability make it ideal for proving out iterative reasoning
• Investment research is harder: context is broad and multimodal, output is qualitative, verification is slow
• The engineering work to put capable models into a research context is underway; the reasoning capability is already there
• Memory today is a convenience shortcut; longer term, continual learning from a firm's history is the unlock that shifts AI from research tool to research driver