Banks reduce AI hallucination risk in finance answers by limiting what the AI can make up. Give it approved definitions instead of raw tables, require every figure to cite its definition and source, have it decline when no approved definition exists, and send answers on key figures through human review before they reach a board deck or a regulator.
Hallucination in finance is often not an invented number from nowhere. It's a real number answering the wrong question.
1. Invented definitions. Asked for core deposits, it builds its own definition from the tables it can see.
2. Wrong source. It picks a table that looks right but isn't the one finance reports from.
3. Fluent gaps. It fills missing context with plausible text, and the answer reads as confident either way.
Ask questions you already know the answer to, including ones with no approved definition. A safe setup gets the first kind right and declines the second.
Preql keeps approved definitions with owners, version history and lineage, and inherits your existing access controls, so AI tools have a governed set of definitions to answer from and reviewers have a record to check against.
Does a better model fix hallucinations?
It reduces some errors. It can't know a definition nobody gave it.
Should AI be used for regulatory figures at all?
With approved definitions, cited sources and human review, it can help prepare them. People should still sign off.

