, APAC
711 views
Photo by freestockcenter via Magnific

Cheap intelligence, scarce judgment: The new model for Asian institutional investing

By Hojun Choi

An investment-native system should build a rich representation of how the investor and the institution thinks. 

Seventy-two percent of buy-side firms across Asia-Pacific are already using artificial intelligence (AI) moderately, according to a 2026 survey of 50 APAC-based investment firms by Bloomberg and WatersTechnology. For Asian investment firms, adoption is no longer an interesting question. The question is what happens to a business built on analytical edge when intelligence becomes cheap.

The first generation of generative AI made existing work faster: Search, summarisation, drafting, and review. The next generation changes what an institution can know, how continuously it can know it, and whether that knowledge outlives the people who produced it.

Over the past year I have sat with heads of research and chief investment officers across New York, London, Singapore, Hong Kong, and Seoul. Two years ago the conversation was about which model was best. This year it is about where an institution’s edge lives once competitors have the same models, and how much of that edge is walking around in people’s heads.

A trade-off that is starting to break
Investing has always faced a trade-off between scale and context. Discretionary investors are good at context: A shift in management language, the second-order effect of a policy change, whether a market narrative makes economic sense. But attention does not scale. Systematic investors solve this by turning information into variables that can be tested across enormous datasets, often stripping away the context that gives them meaning.

That trade-off is particularly important in Asia, where investment decisions span markets with different languages, disclosure conventions and market structures. A system that works on a US 10-K must also be able to interpret a Japanese or Korean filing without losing the context that gives the information meaning.

Large language models (LLM) weaken that trade-off by making qualitative information computable without flattening it. A discretionary hypothesis can be monitored across thousands of companies and tested continuously against contradictory evidence. A statistically discovered relationship can be checked for an economic mechanism before informing a portfolio decision. Discretionary investing does not become systematic, or the reverse. The two architectures meet around a common intelligence layer.

Why this time is different
Sceptics have heard this before. “Quantamental” was the promised synthesis a decade ago, mostly delivering better dashboards rather than a different research process. The reason was technical. Earlier natural-language tools reduced text to a handful of crude variables — sentiment scores, keyword counts, named entities — and discarded the qualifiers that carry meaning. “Revenue” and “North America cloud revenue” collapsed into the same label.

A 2025 study published in Management Science found that large language models can capture information in earnings announcement disclosures by accounting for both individual words and their context, explaining three times more variation in short-window stock returns than dictionary-based and other non-LLM approaches. The study also found that LLM analysis of conference calls further increased explanatory power.

The point is not that language models automatically produce better investment decisions. It is that they can turn qualitative context into testable variables rather than merely summarising text, opening a bridge between forms of analysis that previously sat on opposite sides of the discretionary-systematic divide.

The scarce resource is shifting
Frontier models will keep improving, and access to them will keep broadening. If every large institution can use the same capable models, the model becomes a diminishing source of differentiation. What remains proprietary is everything around it: The institution’s data, investment framework, research, prior decisions and, above all, its judgment.

Much of that judgment lives in notebooks, emails, meeting rooms, and memory. Why an analyst originally liked a company. What would invalidate the thesis. Why a portfolio manager passed on something attractive. What the team learned from a decision that went wrong. When people change coverage or leave, much of that knowledge disappears.

An investment-native system should preserve this context. When new information arrives, it should already know the existing thesis, relevant exposures and prior research, and ask: What changed, which assumption does this touch, what else is exposed, what contradicts our view? Over time it should build a richer representation of how the investor, and eventually the institution, thinks.

This is an operating model, not a tool purchase. Theses become structured records of evidence, assumptions, and invalidation conditions because agents can only monitor what is explicit. Coverage splits: Agents own continuous monitoring, comparison and first-pass synthesis; analysts own the question, causal story, and the call. Post-mortems become inputs rather than documents, so every outcome informs the next decision.

Human judgment does not become less important. It moves up the stack: Defining the right question, challenging causal logic, framing scenarios, sizing uncertainty, making portfolio trade-offs, and owning the risk. The issue is not whether a human is in the loop but where in the loop human judgment creates the most value.

Trust is infrastructure
None of this influences real capital unless the system can be trusted, and trust is a property of the workflow, not a benchmark score. Institutional AI needs provenance, permissioning, point-in-time integrity, auditability, and a clear view of how it fails.

A persuasive answer is worthless if it used information the user was not entitled to see, introduced look-ahead, or cannot show evidence. So evaluate at the workflow level. Did it retrieve the right source? Preserve what the source actually said? Separate evidence from inference? Can the path from data to recommendation be reconstructed? Does it behave the same across different markets and disclosure conventions?

Durable advantage will come not from adopting AI fastest or licensing the most capable model, but from turning proprietary judgment into a system: connected to data, persistent across time, testable, auditable, and able to learn from outcomes. AI’s contribution to investing will not be making the same decisions humans make, faster.

It will be making judgment scalable — and turning what an institution has learned into an asset that compounds.

Join Asian Banking & Finance community
Join Asian Banking & Finance community
Since you're here...

...there are many ways you can work with us to advertise your company and connect to your customers. Our team can help you design and create an advertising campaign, in print and digital, on this website and in print magazine.

We can also organize a real life or digital event for you and find thought leader speakers as well as industry leaders, who could be your potential partners, to join the event. We also run some awards programmes which give you an opportunity to be recognized for your achievements during the year and you can join this as a participant or a sponsor.

Let us help you drive your business forward with a good partnership!