ENDE Find your number1 min

AI & Tech28 Sept 20268 min readupdated 29 Sept 2026

AI and your money: great intern, terrible oracle

What a chatbot can really do for your money in 2026, where it confidently talks crap, and a checklist before you believe a word of it.

I build with AI every day. Which is exactly why I don’t let it pick my investments.

Large language models (the tech behind ChatGPT, Claude, Gemini and friends) are the best intern I’ve ever worked with: fast, tireless, well read. But it’s an intern who never says “I don’t know”. Ever. And when it’s your money, that second part matters more than the first.

What AI is good at#

Everything where you bring the document and the model brings the patience.

  • Explaining documents. Fund fact sheets, the key information document (KID) of a fund, your yearly pension statement, a letter from the tax office. Paste the text, ask what it means and what you have to do by when.
  • Translating bureaucracy. Official letters are a language of their own in every country, spoken by nobody and sent to everybody. A model turns them into plain English, and back.
  • Building and checking spreadsheets. Savings plan, fee comparison, withdrawal plan. Ask for the formulas, then make it explain each one. You stay the reviewer.
  • Categorizing spending. A bank export with a few hundred rows becomes ten categories in a minute. Remove names and account numbers first.
  • Stress-testing a plan. Ask for the strongest arguments against your plan. A model has no ego and no commission. Try getting that at your bank.
  • Summarizing annual reports. Two hundred pages become two. Then you read the parts that matter in the original.

See the pattern? Every time, the facts come from a document you supplied. The model works on your input, not on its memory. That’s where it’s strongest.

Fees are a good first exercise. Let the model explain the cost section of a KID, then run the numbers yourself.

What AI can’t do#

Predict prices#

In its 2025 investor warning, the EU markets regulator ESMA tells consumers that no one and nothing, AI tools included, can accurately predict financial markets, and that there are no shortcuts to wealth. It also points out that public AI tools are not authorized or supervised by financial regulators and have no obligation to act in your best interest.

I don’t agree with regulators often. Here I do.

The same warning covers apps and websites that sell AI “stock signals”, often for expensive monthly or yearly fees. ESMA’s advice: be wary of anyone who claims to predict prices with high accuracy. My version: if the signal worked, nobody would rent it out to you by the month.

And no, this isn’t theory. In May 2026 the German regulator BaFin warned about a website that advertised automated and profitable investment decisions made by AI. According to BaFin, the operators had no licence to offer those services in Germany.

Beat the market on its own#

There is real research. It’s a lot more modest than the headlines.

Pelster and Val (Finance Research Letters, 2024) ran a live experiment with a chat model that had internet access. They found a positive correlation between the model’s ratings and later stock returns, and a strategy based on those ratings produced positive returns. The test was live for a reason: with historical data you can’t rule out that the model already knows the future. The catch is the flip side: the experiment covered a short window around one earnings season. That’s not a track record. That’s a good week.

A 2026 preprint by Crisostomo and Mykhalyuk tested four well-known chat assistants with three prompting strategies. Their finding: the recommendations suffer from recurring reasoning failures, including financial misconceptions and outdated or hallucinated information. With close human guidance the models could outperform the market in their test, and grounding answers in official filings improved accuracy. It’s a preprint, so not yet peer reviewed.

Fun for researchers. No reason to swap a broad index fund for a chat window. And as always, past returns are not promises, whether a human or a model produced them.

Get numbers and tax rules right every time#

A model predicts plausible text. Plausible and correct overlap often. Not always. ESMA warns that AI output can be based on outdated, incorrect or incomplete information.

Tax allowances, contribution rates and pension values change every year, because politicians can’t leave anything alone. A model may quote last year’s figure or mix up two countries, in the same confident tone. In any country, the law and the tax authority are the source, not the chat. For Germany I collected the rules in ETF tax in Germany.

The real danger: scammers with AI#

BaFin has warned repeatedly about ads in which celebrities seem to promise high profits. The people shown know nothing about it, and the money goes to the fraudsters.

Europol’s IOCTA 2026 report calls fraud the fastest-growing area of organized crime, with investment fraud among the most common types. Generative AI lets fraudsters personalize their approach, and voice chatbots pre-screen victims at industrial volume. The crooks got their productivity boost before your bank did.

The EU supervisors suggest a low-tech defence that I really like: agree on a safe word with your family. If “your brother” calls with an urgent request for money, ask for it.

What the law says#

Two things are worth knowing.

First, regulated firms stay responsible. In a public statement of 30 May 2024, ESMA told investment firms that use AI with retail clients that the MiFID II rules still apply, including the duty to act in the client’s best interest. “The algorithm did it” is not a defence. A public chatbot gives you none of that protection.

Second, the EU AI Act sorts AI systems by risk. According to the European Commission, the transparency rules take effect in August 2026: you must be told when you are talking to a machine, and deepfakes must be clearly labelled. Nice. Brussels wrote a rule, and criminals are famous for following rules. So the one who protects your money is still you.

Six prompts that actually help#

Replace the brackets, paste the document text and remove personal data first.

  1. Explain this key information document to me like I am a smart engineer without finance training. List every cost in percent per year and tell me where in the text you found it.
  2. Here is a letter from my tax office. Summarize in plain English: what do they want, what is the deadline, what happens if I do nothing? Quote the original sentence for each point.
  3. Build a spreadsheet formula for a monthly savings plan with [amount], [years], [return] and [fee] per year. Explain each part of the formula. Then give me one test case I can check by hand.
  4. Here is an anonymized bank export as CSV. Sort every row into at most ten categories, sum them per month and list the rows you were unsure about.
  5. Here is my plan: [plan]. Act as a critical reviewer. Give me the five strongest arguments against it, ranked by how much damage each could do.
  6. Summarize this annual report in 300 words. Separate facts from management opinion and give the page number for every figure.

Why do these work? They bring their own source, they demand quotes and page numbers, and they make the model show its uncertainty. And not one of them asks what to buy.

Prompt 3 gives you a formula. Here’s something to check the intern’s result against.

$ compound --monthly 300
€
years
%

Open the full check

Before you trust an AI answer about money#

  • Source: did the answer come from a document I supplied, or from the model’s memory?
  • Numbers: have I checked every figure in the original document or on the official site?
  • Date: is the rule for this year and for my country?
  • Maths: have I recalculated one example by hand or with a calculator?
  • Privacy: did I remove name, address, IBAN, tax ID and anything from an identity document? ESMA advises not to share personal data with public AI tools.
  • Incentive: is someone selling me something on the back of this answer?
  • Reversibility: if the answer is wrong, can I undo the decision? If not, ask a licensed professional.

And yes, this site was made with the help of AI too. The intern is good. I still check its work. Why the site exists is in why I built this site. For the money side, the course covers fees in Fix the leaks.

This article is education, not investment, tax or legal advice.

Sources#

Education, not advice. I don’t know your situation, and past returns promise nothing. Check my numbers, then make your own call. You’re a grown-up.

// keep going

Read next.