Scope and rules

The honest scope

Across five short lessons, you are going to build a bot that helps you research companies, markets, and the news that surrounds them. It will read annual reports and summarise them. It will compare what a company said this year with what it said last year. It will scan news sentiment. It will help you organise what you have read.

It will not place trades. It will not connect to a broker. It will not see your bank balance. It will not tell you what to buy. If you ask it to, it will refuse. That is by design.

This split — research, yes; trading decisions, no — is not because the technology cannot do more. The technology can. The split is because doing more is genuinely dangerous, and because the most useful, most defensible thing AI does in finance is the research half. Trading is where people lose money. Research is where they understand companies better.

Why the trading half is the dangerous half

There is a reason almost every reputable AI-in-finance educator stops at research.

The model does not know what just happened. Language models have a training cutoff. A market that moves in real time on news the model has never seen is a market the model cannot trade. By the time the bot has summarised yesterday's news, the price has already moved.

The model sounds confident when it should not be. If you ask "should I buy this stock", you will get an answer. The answer will be in a tone of confidence the model has no business having. There is no signal in that answer; only generated text that looks like one.

Markets are adversarial. Any "pattern" the model can find in public data has, in expectation, already been priced in by other people running the same model. Edges in markets come from information, interpretation, or risk-tolerance that other participants do not have. An LLM with web access has none of these.

The downside is real money. A wrong research summary is annoying. A wrong trade is a loss you take home. Until something is at the level of "I would let it manage my pension", it should not be the thing making the decision.

None of this means AI is useless in finance. It is genuinely useful — for reading, for summarising, for cross-referencing, for spotting language patterns in earnings calls, for organising news. Those are the things this course teaches.

The legal and regulatory frame, briefly

You do not need to be a lawyer to take this course. But the basic legal picture is worth knowing.

If you are giving someone else advice about investing — for a fee or as part of your job — you are almost certainly subject to financial-services regulation. In the UK, that is the Financial Conduct Authority (FCA); in the US, the Securities and Exchange Commission (SEC) and FINRA; in the EU, MiFID II rules administered by national regulators. These bodies require that anyone offering investment advice is qualified, disclosed, and supervised. A chatbot is none of those things.

If you are using a tool for your own personal decisions, you are not subject to that regulation. But you are also not protected by it. If you act on a hallucinated number from a chatbot and lose money, there is no compensation scheme. There is no firm to sue. The risk is entirely yours.

The EU AI Act, in force since 2024, classifies any AI system used in financial scoring or credit decisions as "high risk", with specific transparency, oversight, and documentation requirements. A consumer-facing research bot used personally is below that threshold today. But the moment such a tool is offered as a service, or used to influence a third party's financial decision, the regulatory burden becomes serious.

The principle to take away: the more your bot is used to make decisions that affect other people's money, the more compliance work you have to do. For learning and personal research, you are fine. For anything else, talk to a regulatory specialist.

What your bot will do

Concretely, by the end of this course, you will have a bot that does the following, in your browser, in Claude or ChatGPT.

• Read a long document — an annual report, a 10-K filing, a press release — and produce a structured summary.
• Compare two documents — last year's report against this year's — and highlight changes in language, strategy, and disclosure.
• Scan a set of news articles about a company and characterise the sentiment, with sources.
• Answer questions about specific facts in a document, citing the location of each fact.
• Refuse, politely, to make a trading or investment recommendation.

That last point is not optional. We will build the refusal into the bot's system prompt in Lesson 2.

What your bot will not do

• It will not predict prices, returns, or market direction.
• It will not tell you whether a stock is "good" or "bad".
• It will not connect to a brokerage account or execute any transaction.
• It will not access real-time market data — only the materials you give it.
• It will not replace a qualified financial adviser. It will not even resemble one.

Your personal rule

One commitment before we move on. Write down — for yourself, somewhere you can find it later — the rule under which you will use the bot you are about to build. Something like:

This is not a formality. The single most common way these tools go wrong is the human user gradually relaxing their guard — first using the bot for research, then for sanity-checking decisions, then for the decisions themselves. The rule you write now is what stops that drift.

Looking ahead

In Lesson 2 we start building. You will write the system prompt — the bot's "constitution" — that defines its role, its constraints, and its mandatory refusal of trading questions. This is the most important single piece of the bot, and we are going to get it right.