The three questions every hedge fund AI announcement should answer

Another week, another hedge fund announcing it is going deeper into AI. The latest is a $20bn London macro fund expanding into equities, with the CIO discussing more granular market views and the head of AI discussing how to find more information in the world. The framing is familiar by now. Advances in technology are revolutionizing what an analyst can do. AI augments rather than replaces. The platform is a powerful engine, still early, with room to grow.

It all sounds right. It is also almost impossible to evaluate.

The problem with most hedge fund AI announcements is that they describe ambition rather than architecture. Allocators, and frankly competitors, who read them learn very little about whether the firm is actually building something that will move returns. Three questions cut through the noise. Any fund serious about AI should be able to answer all three, and the answers reveal far more than the press release ever will.

Question 1: What is the opportunity cost of building rather than partnering?

This is the question almost no firm asks publicly, and it is the most important one.

Every quarter spent building document ingestion, entity resolution, transcript processing, and citation infrastructure is a quarter not spent generating alpha. The largest multi-strategy platforms, the $50bn+ AUM firms that built this stack themselves a decade ago, invested hundreds of millions and a decade of engineering effort. They had the scale, the engineering bench, and the time horizon to justify it. Most firms do not.

The hedge fund industry has long been comfortable buying market data, execution infrastructure, and risk systems from specialists. The AI infrastructure layer is following the same trajectory. The funds moving fastest right now are not the ones building everything in-house. They are the ones treating AI as a composable stack, licensing the data and document infrastructure from specialists and focusing their internal engineering on what is actually proprietary, namely the strategy itself.

A $20bn fund expanding into a new asset class, using an AI platform described as "still in its early stages," is, on any honest read, signaling to the market that it has chosen the slower path. That may be the right call for some firms. For most, the math does not add up.

Question 2: Where does the data come from, and how deep is the coverage?

Most AI announcements lead with the model. The model is the easy part. Frontier LLMs are commoditizing quickly, and the gap between the best and the second-best is narrowing quarter by quarter. The hard part, the part that rarely appears in the press release, is the data substrate underneath.

A discretionary fund expanding from macro into equities needs structured, machine-readable coverage across thousands of individual names. That means primary filings normalized across jurisdictions, transcripts processed with low latency, and, crucially, coverage in the markets where edge actually exists. Global equities are not a US story. China A-shares alone represent more than 5,500 listed companies, most of which file in Mandarin and produce transcripts in Mandarin, and sit entirely outside the major Western data vendors.

Without that substrate, an AI platform is a very expensive way to summarise content that the entire market already has. With it, the same model becomes a coverage multiplier, letting a small team produce institutional-grade research across a universe five times larger than they could cover manually.

So when a fund announces an AI push, the question is not how clever the model is. It is how broad and how deep the data layer is, and whether the firm has access to the markets where information asymmetry still exists.

Question 3: How is every answer audited back to a primary document?

The third question is the one that should keep compliance officers awake.

Generative AI in the investment process is useful only if every output is traceable. An analyst asking a system, "What did this company say about margin pressure on the last call?" needs an answer that includes a citation, a page reference, a timestamp, and, ideally, the surrounding paragraph. Without that, the output is a guess dressed up as research, and no portfolio manager should act on it.

This is where most in-house AI builds quietly struggle. Building a model that generates plausible answers is a weekend project. Building a system that grounds every answer in a specific primary document, with a full audit trail, across millions of filings and transcripts, is a multiyear engineering effort. It is also the difference between a tool a PM will actually use and a demo that lives on an internal wiki.

The honest version of any AI announcement would address this directly. How are hallucinations prevented? What is the citation rate? What happens when the model is asked something that the underlying data cannot support? When those questions go unanswered, allocators should assume the work has not been done.

The quiet shift underneath the announcements

The pattern across the industry is becoming clearer. Funds that announce ambitious in-house AI builds are increasingly those that started later and have the most ground to make up. Funds that are already pulling ahead tend to talk less, partner more, and treat the underlying data and documentation infrastructure as something to license rather than reinvent.

The next wave of hedge fund AI announcements will be more confident, more polished, and more vague. The three questions will not change. Allocators and analysts at these firms should keep asking them.

Talk to Orbit

This is the conversation we have most often with hedge fund and asset manager clients, and it matters deeply to us at Orbit. The data moat, the citation chain, and the deployment model matter more than the model itself.

Orbit is the award-winning AI platform for institutional investment research. We process 70 million documents annually across 75,000+ companies in 120 countries and on 80+ exchanges, including exclusive coverage of 5,500 China A-share names. Every answer is auditable back to the source filing or transcript, and the platform integrates with your team's workflows via Orbit Insight, Cloud APIs, or MCP.

If your firm is weighing the build-versus-partner question, or is already running an in-house AI project and feeling the cost of the data layer, we should talk. Book a conversation with the Orbit team and see what a production-grade research substrate looks like before you commit another quarter of engineering time.