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

Overview of How Investors are using AI — Business Breakdowns (EP.240)

This episode features David Plon, founder of Portrait Analytics, in a conversation with host Matt Russell about practical ways investors are applying AI today. David brings buy‑side experience (barclays trading floor, hedge funds including Balyasny/Baillie?—he describes generalist public- and distressed-equity roles) and explains where AI meaningfully reduces friction in research without replacing conviction-building human work. The discussion covers concrete use cases (idea generation, pre‑buy diligence, and ongoing thesis/position monitoring), prompt and tooling best practices, firm adoption patterns, and a forward view on memory and agentic AI.

Key takeaways

• AI shines at information processing: surfacing relevant signals across broad, noisy data sets rather than replacing core judgment.
• Primary investor use cases today: (1) monitoring portfolio/theses across ecosystems, (2) speeding pre‑buy triage and context building, (3) sourcing ideas aligned to firm mental models.
• Good prompts mimic a clear brief: state the task, why it matters, expected output format, guardrails and domain context.
• Experimentation matters — allocate time to continually test model capabilities; new model releases change the frontier fast.
• For firms, bottoms‑up adoption plus a few firm-level AI services (custom idea screens, thesis monitoring) gets adoption without forcing behavioral change.
• Documenting decision-making (memos, intermediate notes) is high-leverage: it becomes valuable human IP and future ML context/memory.

Tangible use cases and examples

• Position monitoring / ecosystem signals
  • Example: To monitor Expedia, AI can surface hotel industry datapoints (occupancy, pricing, OTA distribution moves) across suppliers, competitors and customers — things an internet/travel analyst might not subscribe to individually.
  • Thesis monitoring: collect soft signals (supplier pricing shifts, customer comments, transcript snippets) that affect an investment thesis in real time.
• Pre-buy research / triage
  • Rapidly kill/advance ideas by surfacing existential risks (management misalignment, egregious comp plans, aggressive accounting).
  • Automate pattern detection: historical guidance behavior (e.g., repeated full‑year guidance revisions leading to “kitchen sink” quarter beats).
  • Move deeper analyses (CEO comp, proxy metric trends) earlier in pipeline — tasks that used to take hours can be a click.
• Idea generation / thematic screens
  • Use AI to find companies exposed to a trend (tariffs, supply-chain geography) including second‑order effects (who has a US-heavy supply chain vs competitors).
  • Encode nuanced mental models (e.g., once-high-quality franchises suffering a temporary hiccup) and search broadly for candidates that fit those qualitative + quantitative patterns.

Prompting & practical LLM techniques

• Treat a prompt like an email to a smart, context‑naïve analyst:
  • Include background/context, the explicit task, why it matters, desired output format, and task guidelines.
  • Provide domain heuristics (e.g., “management is biased positively — take a skeptical view”).
• Iterate: start simple, then add structure and constraints; compare outputs as you increase specificity.
• For structured/precise tasks (financials, pulling numbers, building tables): upload the documents/source data or use a tool that ingests them — this reduces hallucinations.
• For exploratory work (industry map, first-pass narrative): off‑the‑shelf LLMs without full doc uploads are often fine.
• If using raw models, avoid overloading context windows; for large corpora use tools that manage context/chunking.

Workflow, tooling, and adoption at funds

• Best adoption pattern: bottoms‑up experimentation + selective firm-level services that don’t force process change.
  • Useful firm services: bespoke idea-generation screens, centralized thesis-monitoring pipelines that feed analysts passive alerts.
• Encourage individuals to maintain trust: people will only change processes if the tools demonstrably raise conviction.
• Capture firm IP: require / encourage concise memos, meeting notes, and thought-capture. These records are both human-useful and future ML-training/memory fodder.
• Have a “suite of experiments” (10 repeatable tasks) that you run every time a new model is released to recalibrate the capability frontier.

Future outlook — context windows, memory, and agentic AI

• Context windows are larger and models better at using long context, but you should still:
  • Use tools that manage very long corpora instead of manually jamming everything into a chat.
  • Break down overly broad tasks into sub‑tasks if using raw models.
• Memory:
  • Near term: memory features are limited and can be a convenience (store recurring context) rather than transformative.
  • Longer term: persistent, well‑organized firm memories (memos, meeting transcripts, decision history) could let models emulate institutional judgment and pattern recognition — a major competitive edge.
• Agentic AI (models that reason, act, and reflect):
  • Agents are becoming capable in constrained, verifiable domains (software engineering is a leading example).
  • Investment research is messier (heterogeneous sources, qualitative outputs), but agentic workflows are moving from theory to practice; engineering effort is the gating factor.
  • Expect progressive automation of longer-running analyst tasks as tools mature.

Actionable recommendations (quick checklist)

• Start small: use AI to triage new ideas and automate recurring monitoring tasks.
• Build a prompt template: include context, task, desired output, and skeptical domain notes.
• Capture more process output: require short post‑meeting or post‑earnings notes; store memos and idea history.
• Run 10 canonical experiments against new models to understand capability shifts.
• Adopt firm-level passive AI services (custom screens, thesis monitors) to surface ideas without forcing process changes.
• Evaluate vendor tools that handle long context and source ingestion to reduce hallucinations on structured tasks.

Notable quotes

• “The output of an investment research process is ultimately a decision — and hopefully a high-quality decision. AI can reduce friction around information processing without replacing the parts of the process that build conviction.”
• “Where AI has been really helpful is giving you enough baseline context to be able to triage an idea — should I kill this, or spend deep research time on it?”
• “Documenting memos and the thought process is enormously valuable: it’s humanly useful today and will be machine‑useful tomorrow.”

Resources mentioned

• Portrait (Portrait Analytics) — David’s firm and an example of a tool focused on investment research workflows.
• JoinColossus.com — Business Breakdowns host resource for more episodes and weekly summaries.

Summary prepared to help investors and analysts quickly extract practical AI use cases, best practices for prompts and workflows, and strategic steps funds should take now to capture future value.