The first wave of AI in markets was prediction, and it mostly did not work. The second was automation — bots executing rules — which worked in narrow, well-specified cases and failed loudly everywhere else. The wave arriving now is quieter and, for most traders, more consequential: analysis you can talk to.
Four shifts look durable.
From dashboards you read to analysis you interrogate
Every market tool built in the last two decades has the same shape: a screen dense with numbers, and an assumption that you already know which of them matter today. That assumption is what makes professional tools feel hostile to people who are not yet professionals.
The shift underway is from presentation to conversation — not a chatbot glued onto a dashboard, but the dashboard becoming answerable. “Why did the flip move?” “Is this range wide for this ticker?” “What usually happens after a session that looks like this?” The information was always on the screen. What changes is that you no longer have to know which panel to stare at.
Grounding becomes the whole competition
When every product has a competent model behind it — and within a year or two, every product will — the model stops being the differentiator. What separates a useful analyst from a confident one is what it is allowed to see, and what it is required to cite.
Expect the serious tools to converge on a discipline that already exists in research: every number handed to the model rather than recalled by it, every claim traceable to a source, and an explicit refusal when the data is not there. “I do not have that” is a feature. The tools that cannot say it will be the ones quietly inventing statistics — and that failure does not announce itself, because it reads exactly like expertise.
The edge moves to proprietary observation
Public market data is a commodity, and models trained on the public internet all know roughly the same things. So the durable advantage shifts to what a system has observed that nothing else has: its own accumulated record of sessions, positioning and outcomes, built up daily and impossible to acquire retroactively.
That has an unglamorous implication. A tool that started logging two years ago holds something a better-funded competitor cannot buy today, and a tool that started last month simply has to wait. Time in the market applies to the software as well as to the trader.
The line that should not move
The obvious commercial temptation is to let the analysis layer creep toward decisions — from “here is what the map shows” to “here is what you should do” to “here is what I did”. Each step sounds like progress, and each one removes a piece of the judgement that was the reason to have a read at all.
There is a plainer reason to hold the line, too. An analyst reading the market is answering a question that has a knowable answer. An assistant reading your account and telling you what to trade is answering a question about your risk tolerance and your circumstances, and it does not have that information no matter how much of your data it ingests.
What this actually looks like
In practice the near future of AI market analysis is not dramatic. It is a trader opening a dashboard at 9:29, seeing that positioning flipped overnight, asking what that has meant before, getting an answer with the sample size attached and the sources listed — and then making their own decision, faster and better informed than they would have been.
No prediction. No autonomy. Just the distance between data and understanding getting shorter, for more people, every year. That is a smaller promise than the industry likes to make, and it is the one that will hold.