Daloopa CEO Thomas Li says agents now consume nearly all of the company's data, up from about 5% per quarter before AI agents, and year-over-year consumption has grown over 100 times.
Li argues Daloopa's moat is data quality, citing 99.9% accuracy, and says that edge is hard to copy because it rests on process, incentives and culture rather than tools.
Brett Caughran says investment teams are busier than ever and does not expect headcount cuts for at least 18 to 36 months, while Li predicts team structure and analyst hiring will change most.
Agents Now Consume Nearly All Financial Data, Making Quality the Moat
Daloopa argues that in an agent-driven research stack, the value sits in the quality of the underlying data, not in the model or the connector. Agents now use nearly all of the data humans barely touched, so accuracy, completeness and speed matter more. Li says fund workflows, from Excel models to judgment, will follow the same logic. Verifiable facts get automated first, and team structure changes before headcount does.
Takeaways
AI agents now consume nearly all Daloopa's data. Before agents, customers used only about 5% of Daloopa's data each quarter. Agents now consume almost all of it, and year on year consumption has grown over 100 times, which changes what a data vendor is worth.
Data quality and speed decide MCP quality. Li defines an MCP as English prompts in front of an API, so its usefulness depends on the API's data, structure and latency. Top pipes cost a premium, and a single pipe or several makes little difference to a model.
Daloopa's moat is process, not tools. Daloopa claims 99.9 percent accuracy across 6,000 companies and targets completeness, accuracy and speed. Li says rivals cannot copy this easily because it rests on process, incentives and culture, and a good tool does not make a good craftsman.
Facts are verifiable, and judgment may be decomposable. Li says facts and calculations in knowledge work can be checked, and judgment can be learned where it breaks into parts, as principal component analysis does for stock factors. Quant and fundamental funds are racing toward the middle.
Excel modeling needs purpose-built harnessing. Caughran found Claude for Excel and the Daloopa MCP handled simple quarterly updates but not resegmentation. Li says labs lack the detail on how buy side models are wired, so Daloopa is building its own Excel add-in.
Funds build orchestration layers, with access control the trade-off. Funds build internal layers, some using knowledge graphs that pre-compute analysis as markdown files to lift quality and cut cost. Li says skipping the graph mainly helps access control across teams.
Team structure changes first, and headcount holds for now. Li expects the biggest change in how analysts are hired, trained and promoted, with more time on infrastructure and judgment. Caughran sees teams busier than ever and expects no headcount cuts for at least 18 to 36 months.
Open weights versus lab models
Li, speaking personally, is a strong believer in open source and says open weight models are hard to beat on pre training, while labs hold an edge in post training with enterprise customers.
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Caughran argues funds may prefer to post train open models in-house so proprietary knowledge does not leak to foundation labs.
Li answers that post training is more art than science. It works only with a quantifiable objective, as in math proofs, and not for poetry or pitches.