What an AI Trading Assistant Does, and What It Cannot Do

The short answer

An AI trading assistant reads things and writes things. It will summarise a week of news on an instrument, explain what a term means, describe what a chart is doing, talk through a strategy you describe to it, or review a log of your own trades and point at patterns in it.

Those are real capabilities and some of them save genuine time. What an assistant cannot do is know anything about you — your obligations, your tax position, what a loss would actually cost you — and that is precisely the knowledge that separates information from advice.

What it is actually good at

The honest case for an assistant is compression. It takes material you could in principle have gathered and read yourself, and gives it back shorter.

That covers a lot of genuinely tedious work: catching up on an instrument after a week away, getting a plain-English explanation of a mechanism you half understand, having a rule set you have written in prose read back to you with the ambiguities pointed out, or looking over your own trade history for patterns you are too close to notice.

Notice what these have in common. In each case the material already exists, you could have checked it, and a wrong answer is visible to you fairly quickly. That is the zone where an assistant is a decent tool.

What it cannot do, structurally

It does not know your circumstances. Advice is advice because it accounts for the person receiving it. An assistant that has never asked what you earn, what you owe, or how long your money needs to last is not in a position to tell you what is suitable, however personalised its tone.

Confidence is not a signal of correctness. The output is written to read well. A tentative answer and a wrong one look identical, because fluency is what the system produces. This is the failure mode ESMA warns investors about in its own material on using AI for investing — that these tools can generate advice that is inaccurate or misleading, which we cover in more detail in what regulators actually say.

It does not carry the outcome. If it is wrong, the loss is yours alone. Which brings up the part that actually decides how much weight an assistant deserves.

The question that matters: who is operating it?

In May 2024 ESMA published guidance to firms using artificial intelligence in investment services. Its expectation of those firms is stated plainly:

"When using AI, ESMA expects firms to comply with relevant MiFID II requirements, particularly when it comes to organisational aspects, conduct of business, and their regulatory obligation to act in the best interest of the client."

Read that as a statement about accountability rather than about technology. Where an authorised firm puts an AI system in front of you, the firm's existing obligations do not evaporate because a model generated the words.

The same guidance sets out the risks ESMA associates with AI in this setting:

"Algorithmic biases and data quality issues; Opaque decision-making by a firm's staff members; Overreliance on AI by both firms and clients for decision-making; and Privacy and security concerns linked to the collection, storage, and processing of the large amount of data needed by AI systems."

The third is worth dwelling on, because it is the one aimed squarely at the person reading this. Overreliance by clients is named by the regulator as a risk in its own right — not a risk that the model is bad, but a risk in how much weight people place on it.

One limit on what we can tell you here: we looked for a sentence in that guidance explicitly assigning who remains responsible for an AI-assisted decision, and did not find one. The inference is reasonable, but it is an inference, so we are not going to dress it up as a regulator's words.

Using one without being used by it

A few habits make the difference, and none of them require distrusting the tool.

Ask it for its inputs, not its conclusions. "What happened to this instrument this week and when" is checkable. "Should I buy this" is not, and is the question it is least equipped to answer.

Verify anything you would act on. If an assistant cites a figure, a rule or a document, go and look at it. This is the same standard we hold ourselves to in our editorial policy, and it exists because plausible-sounding specifics are exactly what these systems produce most readily.

Keep the decision somewhere else. Write down why you are doing something before you ask the assistant about it. If the reasoning only appeared after the model spoke, you have delegated more than you meant to.

Know which machine you are talking to. An assistant that writes sentences and a model that forecasts prices are different technologies with different failure modes, as what these tools actually are sets out. A tool that does both deserves to be judged as two separate things.

Nothing here is personal advice, and our risk disclosure explains the limits of what this publication offers.