Matt Stockton advises smaller investment firms to name their workflows, systematize one or two, and align the whole team rather than boiling the ocean.
Brett Caughran reports a 1,200,000-token Excel financial model overwhelmed a SharePoint-to-ChatGPT connection, so Stockton recommends splitting tasks, preprocessing documents, and using calculations in code.
Stockton says about 80% of AI work is describing intent clearly, and he recommends recording voice memos so a model can interview him until his process is captured.
Buyside AI succeeds when firms systematize workflows and data first
Moving a buyside firm from individual AI use to a shared team capability is mostly a data and process problem rather than a model problem. Firms must decide which workflows matter, capture where their data lives, and state their intent clearly enough for models to act. Tools change so fast that fluency from hands-on use matters more than any particular build.
Takeaways
Start with workflows and where data lives. Stockton says everything traces to where data lives and how workflows capture it. Many smaller firms lack robust processes, so he advises against boiling the ocean. Name your workflows, pick one or two, systematize them and align the whole team.
Context limits force smaller tasks and preprocessing. Models degrade as they see more, even with million token windows. Caughran measured one Excel model at 1.2m tokens. Stockton splits tasks, preprocesses documents and may use vector databases, as in his project distilling thousand page PDFs. Numbers still belong in structured databases.
Describing intent clearly is most of the work. Stockton estimates about 80% of the work is stating intent. Intuitive investors struggle to write their decisions down, so he records voice memos and has a model interview him. Caughran's team mapped 800 primitive steps and calls cloning your own process a bottleneck.
Fluency comes only from hands-on experimentation. Stockton says no substitute exists for using the tools. Caughran notes that at a 20 person firm two or three people have an AI moment that spreads slowly, and fluency takes hundreds of hours. Many still use AI like a search engine.
Build skills by testing failures and iterating. Stockton advises trying tasks the model will likely fail, then turning good results into reusable skills through hill climbing and reader feedback. LLM-as-judge is tricky. Caughran scores with three LLMs. The failure mode is not reading outputs, and calculations belong in code.
Assemble existing harnesses rather than build custom. Stockton recommends combining harnesses such as Codex or Cowork with company skills and data connectors. Heavy custom builds go stale because newer models need less prescriptive instructions. Needs like scheduled reports can still force real engineering.
Most work does not need frontier models. Stockton argues cost drops sharply when routine work moves off frontier models, citing a Shopify workflow with about a 100x cost improvement. Hy found OpenClaw not worth $50 a day in Opus calls, though he valued its cron jobs.
Data sources and tools mentioned
Investor data sources include transcripts, SEC filings, personal notes and research management systems, and providers now sell model-ready data.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Hy runs a weekly CRM skill over iMessage and Gmail that emails him contacts past their window.
Caughran finds Perplexity Computer more consistent than Claude. It has access to 18 models and spins up sub agents automatically.
Stockton uses Claude Code and Codex daily, never adopted OpenClaw or Hermes, and estimates under 1% of people doing work use Codex.