Brett Caughran says agentic AI moved from hype to useful in late 2025, lifting AI's share of his investor workflows to an estimated 30 to 60 percent.
Brett Caughran won't put LLMs on financial models at about 65 percent accuracy because wrong numbers can look plausible, so he relies on validation agents.
Fundamental Edge advises a nine to eighteen month adoption run, with agents running in parallel as bumpers while its linear process is built out.
Agents make AI useful for investors, but numbers need checking
Investing with AI has moved from chatbots to agents, and the gain is real. Agents run whole workflows end to end, but numbers can still be wrong in ways that look right. The limit for most firms is now how well they can spell out their own process and check the output.
Takeaways
Agents, not chatbots, made AI usable. Agents looked like hype in 2025 until the Claude Code and Claude Cowork moment around December and January. Brett estimates AI covered 5 to 10 percent of his workflows in 2025 and now covers 30 to 60 percent.
Agents work end to end and self-correct. Khe compares agentic workflows to Waymo versus Google Maps, since they carry a task through and fix themselves, such as picking a PDF extraction tool unprompted. Brett cut his prompts from 306 to about 30, then to 18 orchestrated pipelines.
Chatbots still fail on multi-document retrieval. Prompts beyond five or six pages started producing garbled output, and chatbots could not accurately answer a question on Nvidia's ten-year earnings per share history. Brett says ChatGPT's quantitative metrics last year were often pulled from a blog.
Number accuracy is the unsolved problem. Finance benchmarks land around 60 to 70 percent accuracy, and Brett will not put models at about 65 percent on financial models, because wrong numbers look plausible. Validation agents that check numbers and produce spot-check lists are his big unlock.
Investors' imagination limits AI more than technology. Khe says clients anchor to chat and treat prompts like texts rather than briefs. Much knowledge work is vibes in spreadsheets that must be excavated first, and Brett says great portfolio managers run tacit processes, which Fundamental Edge spent four years making linear.
Vendor overpromising fuels demo, disappointment and churn. Brett says vendors claim their tool can do anything an investor can do, and that has caused a cycle of demos, disappointment and churn among buyers.
Outcomes for markets and careers are wide open. Khe expects no single moment when AI is declared ready, only jagged progress, and predicts five steps forward and ten back over the series. Brett says the range of outcomes for markets and Wall Street careers is much wider than usual.
Fundamental Edge's adoption approach
An LLM cannot yet one-shot a complicated three-statement model. It can update a five-year-old model, find key drivers, run research and compare to consensus, which Fundamental Edge breaks into 17 steps.
Brett expects adoption to take nine to eighteen months, with agents running in parallel to the existing process as bumpers.
Tap a name to go there · tap a line to see why they are linked