Spotting AI mistakes in finance before they cost you

A confident answer is not necessarily a correct one. In finance, where a small error in a figure can change a conclusion, it helps to know where AI tends to go wrong and how to catch mistakes before they affect your decisions. This article lists the common failure modes, with examples, and gives practical checks you can apply every time.

Why mistakes happen

AI language models generate text that is likely to follow from what came before. They do not look up facts in the way a database does, unless they are connected to a data source, and even then they can misread or misreport what they retrieve. The result is output that sounds authoritative whether or not it is right. This is sometimes called hallucination: a plausible but invented detail.

Common failure modes

• Outdated information. The assistant describes a company's situation as of an earlier date, such as an old share price, a former chief executive or last year's guidance, as if it were current.
• Confusing similar entities. Companies with similar names, different share classes of the same company, a parent and a subsidiary, or listings in different countries.
• Mixing up periods. A quarter versus a full year, calendar versus fiscal years, or trailing versus forward figures.
• Unit and currency errors. Thousands versus millions versus billions, or figures in different currencies presented as if they were the same.
• Arithmetic and logic slips. Multi-step calculations such as growth rates, margins and compound returns can go wrong, especially if done in a single pass.
• Invented details. A quotation, a statistic, a source or a filing section that does not exist. If a citation looks specific but you cannot find it, treat it as suspect.
• Overconfident summaries. Glossing over uncertainty, or presenting an estimate as a fact.
• Misreading a table or a chart. Particularly when documents are long or poorly formatted.
• Agreeing with you. Tending to endorse the premise of your question, even if it is flawed.

Simple checks that catch most errors

1. Check the date and period of every number. Is it the latest filing? Which quarter or fiscal year? Trailing or forward?
2. Confirm key figures in the original source. Use the company's own filing or a trusted data provider for any number that matters to your decision.
3. Recalculate important ratios yourself. A growth rate or margin takes a minute to check with a calculator or a spreadsheet.
4. Check units and currency. Look at the table headers in the filing.
5. Look for the citation. Ask where each figure comes from, and then open the source. If the assistant cannot name one, do not rely on it.
6. Test consistency. Ask the same question in a different way, or ask the assistant to double-check its figures. Different answers are a red flag.
7. Compare with a second source, such as another data provider or the previous quarter's report.
8. Ask for uncertainty. "Which parts of your answer are you least sure about?" can surface weak spots.

A worked example

Suppose an assistant tells you a company's revenue grew 30% last year. To check:

• Open the income statement and find revenue for the last two fiscal years. Say they are 1,000 and 1,200.
• Growth is (1,200 minus 1,000) divided by 1,000, which is 20%, not 30%.
• Possible explanations: the assistant used a different period, included an acquisition, used pro forma figures or simply got it wrong. You now know to look further.

A minute of checking has changed the conclusion.

Higher-risk situations

Be extra careful when:

• The decision involves a large amount of money.
• The figures are the main support for your thesis.
• The company is small, obscure, or recently changed its name or structure.
• The information is very recent, such as news from the last few days.
• The assistant cites a regulation, a tax rule or a legal point. These vary by country and change.

Keep the decision yours

AI is excellent for speed and breadth. Use it to widen your research and save time, while keeping responsibility for verifying the facts and for the choices you make with your money. If you notice a mistake, treat it as information about how much to trust the rest of that answer.

Common mistakes

• Accepting a number because it is precise. A figure with many decimal places can still be wrong.
• Treating a fluent explanation as proof.
• Checking only the surprising claims and trusting the rest.
• Using AI output without noting its date.

A short checklist

• I checked the date and period of each key figure.
• I confirmed important numbers in the source.
• I recalculated the main ratios.
• I looked for invented citations.
• I decided for myself.

Spotting mistakes is a habit, not a talent. A few simple checks, applied consistently, catch most problems before they cost you anything.