What AI Changes About Trading Research, and What It Leaves to You

The short answer

AI has made the gathering stage of trading research much cheaper. It reads more than you will, summarises it quickly, and runs the same checks on a bad day as on a good one. That is a real gain and it is worth using.

What it has not done is make the understanding cheaper. The tools state wrong things in the same assured tone as right ones, they know nothing about your circumstances, and the institutions supervising them most closely are candid about only partly understanding them. So the sequence that works is unchanged: use them for breadth, check anything you would act on, and keep the decision.

Why this arrived so quickly

The sense that AI turned up in trading all at once is not an illusion, and it is measurable further up the chain. In the Bank of England and FCA's 2024 survey of UK financial services, 75% of the 118 firms that responded were already using AI, with another 10% expecting to within three years. A third of all the use cases they described were third-party implementations — bought, not built — and that share had roughly doubled since the 2022 survey.

Retail tooling follows institutional tooling, usually at a distance and usually in a cheaper form. What reaches a private trader now is the same basic capability: something that reads a great deal quickly and writes a readable summary of it. It is worth knowing what is actually inside these tools before deciding how much of your process to hand over to one.

That is worth taking seriously. It is also worth being precise about, because the gap between what these tools do well and what they are sold as doing is where the money goes.

What the tools are genuinely good at

Strip out the marketing and the useful capability is compression. A model reads more than you will and condenses it. Applied to research, that shows up in a few specific places:

• Breadth on a schedule. Checking what is scheduled across several instruments, rather than the one you already like.
• Summarising the unfamiliar. A sector, an instrument or a policy decision you have not followed, reduced to something you can start asking questions about.
• Restating your own plan. Having an idea described back to you exposes the parts you had not actually thought through.
• Arguing the other side. A stated case against the trade you are leaning towards is a cheap check on your own reasoning, and the tools are good at producing one on demand.
• Consistency. The same checks run on a tired Friday as on a fresh Monday. People do not manage this; software does it by default.

None of that is trivial. Research is mostly bookkeeping, bookkeeping is tedious, and tedium is where people quietly stop doing the work.

The gap nobody advertises

Here is the finding from that same survey that deserves more attention than the headline adoption number: 46% of firms reported only "partial understanding" of the AI technologies they use, against 34% claiming complete understanding. The report attributes the gap mainly to third-party models — the ones firms buy rather than build.

Sit with that. These are regulated institutions, with compliance functions, model risk teams and a legal obligation to supervise what they deploy. Just under half say they only partly understand the thing they are running.

A private trader using a consumer product sits further from the model than any of them. There is no version of this where the right posture is deference.

Why the decision stays with you

ESMA's May 2024 statement to firms using AI in investment services is short and worth reading in full. Two things in it carry over to anyone using these tools privately.

The first is that it treats AI as a tool used within existing obligations rather than as a new category that suspends them — firms are expected to comply with MiFID II requirements, including the duty to act in the client's best interest, whatever software they put in front of people.

The second is the risk list, which names overreliance on AI by both firms and clients as one of the risks ESMA associates with these systems, alongside algorithmic bias, data quality and opaque decision-making.

Overreliance is the realistic failure mode. It is not a dramatic error. It is the slow substitution of a summary for your own reading, continuing right up until the summary is wrong about something that matters — and a wrong answer arrives in exactly the same assured tone as a right one. There is no change of register to warn you.

How much judgement to delegate is the open question underneath all of this, and we have looked at the evidence on that trade-off separately. It does not resolve into a rule, which is itself the point.

A research routine that survives a wrong answer

The way to use something unreliable is to build the unreliability into the process rather than hope it does not show up.

1. Ask for description, not instruction. "What has this level done before, and what would invalidate this idea?" is answerable. "Should I buy?" is not, and the answer will sound just as confident.
2. Check anything you would act on. Every figure, date and contract specification, against the source — the exchange, the broker's own documentation, the official release. If a claim cannot be checked, it cannot be traded on.
3. Keep sizing out of it entirely. Position size is arithmetic from your stop distance and your risk limit. It is not a thing to ask an opinion about.
4. Write down what would change your mind, before you are in the position rather than after.
5. Review the decision, not the outcome. A good decision can lose and a bad one can win. Grading yourself on results alone teaches the wrong lesson.

Step two is the one that gets skipped, and it is the one that makes the other four worth doing.

What to ask of anything that teaches with AI

A fair amount of trading education now has an AI component bolted on, and the useful question is not how advanced it is. It is whether the thing is built to make you better at judging, or to make judging feel unnecessary.

Three questions separate them, and they are the same ones worth putting to any tool before you rely on it:

• Does it show its reasoning, or just its conclusion? A level with an explanation can be argued with. A signal with nothing attached cannot be evaluated or sized.
• Does it say what would make it wrong? Anything that only describes the case for a trade is marketing, whoever or whatever wrote it.
• Is there a person to ask? The questions that matter while learning are usually the ones you do not yet know how to phrase, which is precisely what a search box cannot help with.

The honest version of AI in trading education is unglamorous: it does the reading, keeps the preparation consistent, and leaves the judgement — and the consequences — where they were. That is a real gain. It is just a smaller and more specific one than the category usually claims for itself.

The short version

AI changed the cost of gathering information. It did not change what you have to understand before acting on it, and it did not move the responsibility anywhere. The institutions supervising this most closely are the ones saying so most plainly: just under half admit they only partly understand their own systems, and their regulator lists overreliance as a named risk rather than a hypothetical one.

Use the tools for breadth. Keep the verification step. Keep the decision.