Summary

Bustamante’s highest-leverage recommendation is to break the investment process into explicit tasks, then decide “which one can I delegate to AI?” A prior-quarter memo can become the template for the next earnings update, while screening can combine ordinary valuation filters with transcript-level conditions: founder leadership, the CEO mentioning future buybacks, cash available for repurchases, and a record of buying below intrinsic value.
AI’s current investor edge is retrieval and synthesis at a scale that turns a day of work into seconds or minutes. Walker’s examples include searching roughly 50 Caesars transcripts for acquisition commentary and assembling a decade of restaurant same-store sales; Bustamante extends the ladder from one-company KPI extraction to peer comparisons and whole-market qualitative-plus-quantitative scans. The next breakthroughs are “offline and parallelization”: hundreds of simultaneous questions and research that continues overnight.
The durable human role shifts from producing analysis to architecting, challenging, and judging it. Bustamante says AI went from writing 5–10% of his team’s code to effectively 100%, and expects finance to follow: an analyst covering 50 names might cover 100 while doing more complex work. The scarce skills become “taste,” pattern recognition, workflow design, and responsibility for the final bet—not mechanical summarization.
Concentrated investing remains resistant because markets are fat-tailed and the winning edge case can look indistinguishable from a deserved zero. Carvana carried leverage, deteriorating fundamentals, governance concerns, related-party dealings, and short research; Tesla could be framed as a carmaker worth more than the global auto industry or as an Elon Musk option on entirely different businesses. “That’s why we need the human in the loop”: a base-rate model can rank a basket, but a ten-stock fund must choose the exceptional survivor.
Trust depends on constrained sources, citations, and verification rather than a model’s fluency. Bustamante calls SEC filings the “ultimate source of truth,” cites Fintool at roughly 98% on FinanceBench versus about 40% for ChatGPT and 50-something for Perplexity, and says broader web retrieval must verify each fact before incorporating it. One hallucination or wrong diluted-share count can make a PM “lose trust.”
Delegating extraction need not eliminate learning if the investor deliberately moves manual attention up the stack. Walker’s compromise is to let AI summarize years of executive compensation and peers, then read the latest proxy himself; in one Chipotle analysis, cross-company comparison would surface a buried provision paying compensation in RSUs when performance reached 200% of target. AI can also monitor changing 14A metrics and potentially off-cycle Form 4 grants, though Walker’s experiment produced only four or five valid cases from 20 Fintool results.
Bustamante sees a large current adoption gap, while Walker fears AI will become table stakes. Some large firms still prohibit ChatGPT, while smaller Kennedy Capital in St. Louis is described as deeply AI-enabled. Walker’s darker conclusion is that non-users may “get your face ripped off,” while users merely keep pace as markets become more efficient.

Deep dive

1. Workflow decomposition turns a chatbot into an analyst

Bustamante’s opening prescription is procedural: map the path from idea generation to investment—or rejection—into perhaps 50 specific tasks, then ask, “Which one can I delegate to AI?” Funds often already possess this map because they use it to train junior analysts.
For a quarterly Home Depot memo, the system can ingest the Q3 example, read the Q4 earnings call, 10-Q, press release, and 8-K, then reproduce the established structure and extract the required figures. On the sell side, the analyst reviews it, adds the summary or conclusion, maintains or changes the price target, and publishes quickly.
Hedge-fund memos are less standardized: the relevant template may vary by company, industry, PM, and thesis. A margin-pressure thesis might demand inflation mentions and specific operating indicators, so the valuable input is not a universal memo but the investor’s own prior example.
Screening shows why task definition matters. A conventional filter might specify market cap below $10 billion and P/E below 30; AI can add founder leadership, management’s buyback language, balance-sheet capacity, and whether previous repurchases occurred below intrinsic value, then rank the resulting opportunities.

2. Personal pattern matching requires more than a small portfolio

Walker has struggled to upload a successful thesis and ask for similar ideas, or compare today’s portfolio with his historical winners and losers. Bustamante agrees this is “super complicated”: Fintool reserves it for large enterprise work requiring portfolios, coverage universes, internal memos, stock-price data, and SharePoint information.
The sparse-data objection is sharp. A concentrated manager holding eight stocks for two years makes roughly one new investment per quarter, unlike Renaissance’s enormous stream of trades; there may be too few observations for AI to distinguish genuine skill from an idiosyncratic winner.
Bustamante’s counterclaim is categorical: with enough SEC filings, calls, presentations, expert-network transcripts, internal criteria, and contextual data, language models will eventually search qualitative opportunities just as machines conquered quantitative ones. “I’m 100% sure it will happen,” although the market should become more efficient and alpha harder.
Fintool itself grew from Bustamante hearing Buffett describe scanning small caps and reading a booklet of Japanese companies to find what made business and management sense. His premise was that this previously human, partly qualitative assessment had become machine-readable.

3. The analyst’s job moves from production to orchestration

Bustamante’s software analogy is deliberately provocative: AI progressed from writing 5–10% of his team’s buggy code, to 30%, then 50%, and now effectively 100%. His conclusion—“AI is the best software engineer on the planet”—supports a finance role modeled on the architect or “meta thinker.”
At a large bank, juniors who summarize 10-Ks in standardized language are directly exposed. The constructive version is that an analyst covering 50 names can cover 100, or surround Chipotle with five relevant peers and produce a richer study rather than another basic note on inflation.
Walker wonders whether embodied “gumshoe” work becomes the new edge: attend a franchisee meeting, sense morale at three out of ten, and compare that observation with the seven-out-of-ten expectations embedded in the stock. Bustamante concedes proprietary data always helps, while anticipating systems that eventually analyze management videos and patterns of excitement or concern.
Public qualitative data remains underused. Bustamante says that at FinChat they downloaded long-form podcasts and asked AI to isolate investor-relevant discussion of capex, product launches, competitors, and whether LLMs are becoming commodities—information a tech PM cannot manually extract from every four- or six-hour appearance.

4. Parallel and offline agents could invert the research workflow

Today’s capability ladder begins with a sourced table—such as stores opened in each of the past eight quarters—then expands to five-company KPI comparison, management commentary, and whole-market scans blending adjusted financial metrics with qualitative criteria. Bustamante says the largest queries can run 20–50 minutes.
Walker calls his own usage “super Google”: Fintool searched years of Caesars acquisition commentary in 30 seconds rather than a day, while AI can rapidly assemble same-store-sales histories for Wendy’s, McDonald’s, and Burger King. Useful as that is, he asks how it makes him smarter rather than merely faster.
Bustamante’s first answer is parallelization. Instead of asking five primer questions sequentially, an investor could request 100 or 200 analyses at once: business model, how the company makes money, financials, valuation, CEO compensation, and peer benchmarking.
The larger inversion is offline work. Once an agent understands that an investor values cash-rich companies, proven capital allocators, possible spin-offs, or CEOs arriving from competitors, it could research from 7 p.m. to 9 a.m., score moat and management, push candidates proactively, and learn from the investor’s feedback.

5. Long-horizon AI backtests are contaminated by hindsight

Walker asks whether 20 years of successful investments create a moat over a 21-year-old beginner. Bustamante notes that anyone can study historical winners such as Apple or Coca-Cola, but Walker’s Philip Morris challenge exposes the problem: its brand, addiction, distribution, and low P/E looked attractive, yet government action could plausibly have destroyed the equity.
Walker argues that a useful qualitative system would need contemporaneous context, not just outcomes: the news, prevailing narrative, policy statements, expert views, and what investors could have known then. Only with that surrounding corpus might it judge whether tariff panic reflects genuine impairment or “Mr. Market” becoming irrational.
Backtesting remains structurally weak. An eight-position portfolio held for three years yields only about 51 observations over 20 years, while a modern LLM already knows what happened after any historical cutoff; prompting it to ignore Apple’s future does not remove that knowledge. Bustamante sees short-horizon news trading as more tractable than proving a ten-year concentrated strategy.

6. Delegation should move learning up the stack

Walker’s strongest objection is that building a model manually creates understanding: entering 5% revenue growth and tracing it through the statements teaches something a finished spreadsheet cannot. Bustamante accepts the risk—“the more we do the manual work, then we learn”—but argues extraction is no longer the right learning frontier.
Their compromise is layered work. Let AI summarize five years of compensation and benchmark peers, then read the focal company’s latest proxy manually; the investor arrives with historical context while still wrestling directly with the document that governs the present thesis.
In the Chipotle example, a customer manually unpacked a roughly $2 million cash incentive, with 75% tied to company-performance factors and 40% of that linked to measures including comparable restaurant sales and cash-flow margin. Bustamante says a five-CEO comparison would flag a missed footnote: if achievement reaches 200% of target, the remaining compensation is paid in RSUs rather than cash.
Bustamante describes an analysis that scans successive 14A filings for changed incentive metrics—say, net income falling from 50% to 10% of the formula—and tests whether changes coincide with deteriorating results. Walker adds off-cycle Form 4 grants as a possible signal of good news ahead, while acknowledging current extraction is noisy.

7. Fat tails preserve the need for human judgment

A Fintool presentation on Carvana surfaced the bearish evidence easily: leverage, deteriorating performance, a related company run by the founder’s father lending money to Carvana, and the Hindenburg report alongside other short research. Walker says that fact pattern would make the base case a zero.
Bustamante’s answer is uncertainty, not retrospective certainty: Carvana “might have been a zero,” might still prove one, yet the rising stock could devastate a short. Finance is “fat tail,” and when being wrong has consequences, “in one second you can be wrong and it’s game over.”
A quant can short a diversified ETF of AI-identified red flags; a concentrated long-only investor with ten positions cannot rely on the basket outcome. Bustamante therefore keeps a highly paid human responsible for the insight and the bet, even if AI performs virtually all preparatory analysis.
Tesla makes the narrative collision explicit. Walker contrasts a roughly $300 billion global auto industry with Tesla around $1 trillion, while the bull case is “Elon Musk is taking us to the moon”; Bustamante says valuation also requires judgment about Musk, regulatory credits, Optimus robots, and a proposed NVIDIA competitor. The stories cannot be reconciled by arithmetic alone.

8. Source control is the prerequisite for investor trust

Bustamante says one large customer initially dismissed Fintool in favor of ChatGPT, then realized after about three months that it produced nonstop hallucinations. On FinanceBench, he cites approximate accuracy of 98% for Fintool, 40% for ChatGPT with search, and 50-something for Perplexity; these are his stated figures, not independently examined in the conversation.
Fintool’s sequence was to master filings as the “ultimate source of truth,” then add investor presentations, earnings calls, and YouTube transcripts. Web search was due within two weeks, but each retrieved fact would require verification before entering an answer.
The commercial standard is unforgiving: returning a hallucination or wrong diluted-share count is “the best way to lose a customer.” That explains why finance retrieval products can feel slow—the answer passes through multiple sourcing and validation steps before a PM sees it.
Uploads are most valuable when the data is genuinely private: internal memos, Excel files, sell-side research, expert calls, or recorded management conversations. Public work can still focus the model; Bustamante cites SemiAnalysis on liquid-cooled data centers and an Abilene site expanding from 100,000 toward 200,000 GPUs as context for prospective GPU demand.

9. Better inputs—and sustained curiosity—remain human advantages

Bias does not disappear when private data arrives. Walker found ChatGPT favored the legal side represented by six uploaded filings over the side represented by five; management transcripts similarly skew bullish because CEOs emphasize transformations and cost cuts. Bustamante’s remedy is to privilege numbers, compare several calls, and benchmark peers rather than score executive language naively.
For recorded calls, Bustamante points to Granola as a leading option that can preserve the raw transcript and create notes. He says extracting additional information ultimately comes from asking small questions.
The best questioning example came from a Datadog investor-relations executive: ask a company about its competitor—Dollar Tree about Dollar General, or McDonald’s about Burger King—then infer what the answer reveals about inflation, inventory, and management’s own preoccupations. Bustamante’s conclusion is that asking good questions remains an art.
The closing disagreement is about how long the edge lasts. Bustamante says tools can change completely in three months and urges investors to “stay curious”; some major firms still ban ChatGPT while Kennedy Capital in St. Louis is already highly AI-enabled. Walker fears AI will become “table stakes”: laggards lose badly, while users keep pace as markets become more efficient and returns compress.

After Interviewing AI Founders, Here's What You Really Need to Know | EP #175 No Priors Ep. 116 | With Sarah and Elad