It’s easy to be swept away by the hype surrounding generalist tools like ChatGPT. But as anyone working in investment management who has tried these tools knows, real-world research and decision-making demand precision and consistency that generalist tools cannot deliver. At Hudson Labs, we built the Co-Analyst because investment professionals require more than generic answers—they need structured, timely, and precise data.

‍

I recently joined Ehsaan Ehsani, Executive Director at Crescendo Partners and the author of the newsletter, Value Hedgehog, for a conversation on generalist AI and value investing, and why specialized AI like the Hudson Labs Co-Analyst is not only useful but necessary in financial research. The following is a summary of the main talking points. You can watch the entire conversation here.

The Co-Analyst Is Not Just a Model—It's a System

I’m often asked how the Co-Analyst compares to ChatGPT. The question itself reflects a misunderstanding. ChatGPT is a product built around a generalist language model. The Co-Analyst is a system that includes multiple fine-tuned models, real-time data retrieval, reasoning pipelines, and domain-specific logic.

‍

What we’ve built at Hudson Labs isn’t simply a better AI model. We built a modelling architecture that facilitates financial retrieval, understands guidance, and supports real-time analysis.

Why Retrieval Is Fundamental to AI Architectures

Language models have a context window, which limits how much information they can process at once. Think of this as the prompt size when you’re copy-pasting something into ChatGPT. This creates a major obstacle in finance, where the relevant detail may be buried on page 42 of a filing released yesterday.

‍

If the model hasn’t seen the document before, it must access that data in real time. We solve this with retrieval-augmented generation (RAG). We pre-process documents, tag them by topic, and store them in a way that makes retrieval accurate and efficient. Our models know when figures are in pesos versus dollars, and they can align context from multiple parts of a document or from multiple sources.

Why General Models Fail in Finance

Generalist models like GPT-4 often impress in demos, but when applied to new, domain-specific tasks, their performance drops drastically. We ran out-of-sample tests with original financial reasoning questions and GPT-4 answered just 12% of them correctly. That’s because it wasn’t reasoning—it was repeating memorized content.

For investment workflows, that’s just not good enough. A misinterpreted number or a missed sentence about future guidance can materially affect an analyst’s target for the stock. Our clients rely on us for accuracy, not close approximations.

Specialization Drives Performance

The Co-Analyst handles tasks that generalist models cannot. When a user asks for guidance, outlooks, or numerical trends, we route those queries through specialized pipelines. These models don’t just look for keywords—they extract and validate the exact figures, apply arithmetic checks, and present structured outputs.

‍

We also perform boilerplate suppression. Investment documents are full of legal noise. We strip that out and focus on what matters. For example, the phrase “we could be subject to an SEC investigation” means little. “We are subject to an SEC investigation” means a lot. Our systems know the difference.

‍

We structure and rank this data before the model sees it. This helps the model focus on high-impact, relevant content, reducing error and noise.

A Quick Demonstration

When I ask the Co-Analyst for CapEx guidance at Meta over the last eight quarters, it extracts the data, converts it into a time series, and presents it with sources—within seconds. Generalist models tend to confuse actuals with guidance or miss quarters entirely.

‍

Another example: when analyzing Costco’s traffic trends, the Co-Analyst pulls that data from earnings calls, even though it’s not in standardized tables. It formats everything for immediate use—no copy-pasting individual numbers, no manual cleanup.

‍

We also address arithmetic failures. Generalist models often struggle with simple math. The Co-Analyst performs checks in real time. If a company gives guidance of $37.5B ±2%, the Co-Analyst computes the bounds and presents them as usable figures—not vague estimates.

‍

Final Thoughts

The investment community has been cautious with AI, and for good reason. Many tools overpromise and underdeliver. Wrapping a generalist model around SEC EDGAR and calling it an investment research platform doesn’t work.

‍

At Hudson Labs, we’ve taken a different approach. We built an architecture that meets the standards our users demand. nalysts, PMs, and researchers cannot afford ambiguity. The Co-Analyst doesn’t replace expertise; it accelerates and sharpens it.

‍

If you’re building or evaluating AI tools in finance, ask this: Does the system retrieve, reason, and respond based on your needs, or is it just guessing based on what it has seen before? The difference is profound.

‍

To learn more or see the Co-Analyst in action, book a demo here. We’re always open to showing you how specialized AI can raise the standard.