Blog / 16 Nov 2026

If AI can learn investment judgement, what’s your edge?

Judgement is learned understanding. AI can learn it too, and copying how humans invest may set the ambition too low. Where does that leave investment edge?

“AI won’t replace my judgement.” That’s a big assumption to build an investment career around.

Investors often talk about judgement as the part of their job AI won’t be able to do, but if much of that judgement comes from experience and learning, why couldn’t a machine learn it too?

I think AI can learn most, if not all of what makes an experienced investor able to generate consistent high risk adjusted returns.

Arguably more importantly, I don't think we should just focus on being able to replicate the ways humans currently invest and frankly, I think it sets the ambition too low.

Human judgement is limited by what we’ve seen, what we remember and how well we learn from our mistakes. There may be better ways to do it.

If AI can both learn our judgement and improve on it, what does your investment edge look like?

What is judgement?

Judgement is applying what you’ve learned about the world to an uncertain situation. That learning comes from experience, study, other people and reasoning.

Over time, it forms mental models that shape what you notice, which explanations you consider and how much weight you give the evidence, which is why two investors can observe the same thing and reach different conclusions (which makes a market).

We all know that investing is a probabilistic art, where we filter new information through those mental models to update the range of possible outcomes and the likelihood we assign to each. Good judgement means knowing how much the evidence should change our view, including when it shouldn’t change it at all.

Kahneman makes this point about intuition in Thinking, Fast and Slow, the book every investor has read (or at least owns). Expert intuition comes from recognising learned patterns, but developing it requires reliable patterns and useful feedback. Years of experience can build good judgement, or build a lot of false confidence.

Our mental models therefore contain both understanding and mistakes. Improving our judgement means examining how we reached a conclusion and revising the underlying mental model when the evidence calls for it.

How AI can learn investment judgement

Our work on systematic alternative data and mosaic theory at scale started with human knowledge. We took methods an analyst would normally acquire through years of experience and encoded them as research skills: how to interpret a dataset, connect evidence across businesses, and investigate why a relationship breaks down.

The models then successfully applied those lessons to new evidence at scale. We supplied the understanding; the systems extended how widely it could be applied.

Our regime-shift research explored a totally different approach. We had AI agents study 132 historical cases of regime shifts and false alarms in businesses, extract recurring mechanisms and warning signs, and turn those findings into skills for other agents to apply to unfamiliar companies. The system helped build the understanding it would subsequently use.

This goes beyond transferring an experienced analyst’s knowledge into software. It creates a way to build experience from historical evidence, with a breadth that isn’t limited by one person’s career, coverage, or memory.

We can scale the process of acquiring understanding, not just the process of applying it.

Can AI find better ways to make investment decisions?

Human investment methods were developed around strict natural limits on how much we can read, remember, and investigate. A system without those limitations will find better ways to do the work. Teaching AI to reason like an experienced investor is a good starting point, but it's not the final form.

Rich Sutton’s essay The Bitter Lesson gives this thought process historical weight. He argues that general methods which let machines search and learn have consistently outperformed approaches built around human knowledge as computation grows.

Applied to investing, that suggests giving systems room to discover useful signals and research methods we haven’t thought to teach them.

One possibility is a system learning how to improve its own judgement. If a model repeatedly overweights a certain kind of news, a research system could investigate the error, test a correction, and retain it if subsequent forecasts improve. It could do similar work on which sources to trust, or identify exactly when a forecasting method stops being useful.

That improvement would still need to be proven on new outcomes. But if a system can learn how to make decisions, and then learn how to improve the process behind them, we shouldn't assume human capability sets the limit.

So, what does an edge look like going forward?

Simply becoming a better analyst using current methods of investment research is unlikely to be enough. As AI learns to do more of that work, investors will need to develop methods that go beyond what a human team could previously achieve.

The most important thing to understand is that edge is never a constant. Human investors have always had to evolve their methods to stay ahead and that doesn't change in a systems-driven market.

For example, investors might suggest looking for places where AI models make predictable mistakes. That can work in the short term, provided the mistake actually distorts prices enough to pay off after trading costs and risk.

But you cannot build a lasting strategy on static blind spots. The moment a recurring error becomes predictable, competing systems will learn to fix it. Every useful discovery eventually gets copied, priced in, and commoditised.

When insights decay that quickly, retaining your edge becomes an engineering problem.

You can no longer rely on a handful of brilliant insights, datasets or a single static model. Your advantage comes from the architecture of your research engine: how fast it can spot when an old signal stops working, test new hypotheses, and upgrade its own process without human bottlenecks.

That is why we are dedicating most of our time to building these systems @primerapp_.

We are exploring what becomes possible when a machine can investigate far more evidence, test more explanations, and retain more of what it learns than any human team ever could.

Building that learning capacity, including the capacity to discover approaches we haven’t thought of ourselves, will become the defining advantage in public market investing.