Why AI gets your financial numbers wrong

AI tools often get financial numbers wrong because they don't know your definitions. Asked for "deposits" or "revenue," a model picks a table and a calculation that look reasonable, with no way to know which one your CFO approved. The answer sounds confident either way. Giving AI access to agreed definitions, not just raw data, makes its answers far more consistent and easier to check.

Many AI errors in finance aren't arithmetic mistakes. They're interpretation mistakes. The model added the numbers correctly, but it added the wrong ones.

Where AI goes wrong

1. It guesses the definition. "Commercial deposits" could mean three different things across your core, GL and FP&A model. The model picks one without saying so.
2. It can't see the adjustments. Finance numbers include accruals, eliminations and manual entries that live outside the source tables.
3. It doesn't know which source wins. When the CRM and the GL disagree, a person knows which one to trust for which question. A model doesn't.
4. It sounds sure regardless. A wrong number in a fluent paragraph is harder to catch than a wrong cell in a spreadsheet.

A bank example

Ask an assistant for commercial deposit growth last quarter. It finds a deposits table, filters on a product type field, and returns a figure. The finance team's figure excludes treasury sweep balances and uses average, not ending, balances. Both are reasonable calculations. Only one matches the number in the board deck.

What AI needs to get it right

1. Definitions. What each metric means, written down and approved.
2. Mappings. Where each concept lives in each system.
3. Rules. Which source wins for which question.
4. Lineage. Where every figure came from, so a person can check it.

How Preql helps

Preql builds a context layer of your institution's approved definitions, mappings and lineage, so AI tools can work from the same definitions your finance team uses.

FAQ

Can I trust ChatGPT with our financials?

Not on its own. Without your definitions, it is guessing at what your numbers mean.

Can I trust Claude with our financials?

Not on its own. Without your definitions, it is guessing at what your numbers mean.

Does a better model fix this?

A better model reasons better. It still can't know a definition nobody gave it.