AI Hallucinations in Banking: The Risk Is Embarrassment
AI hallucinations in banking are usually framed as a security problem. For the people who use AI at the top of a bank, the bigger risk is personal: an AI-made figure, repeated in a committee, that turns out to be wrong.
Data theft is a policy question that IT and compliance own. Embarrassment lands on one person, in one meeting, in front of colleagues whose trust took years to earn. That is why senior bankers stop using a tool after one bad number, long before any breach.
Why does one wrong number end the relationship?
A board pack is built on trust that flows upwards. The analyst trusts the source, the director trusts the analyst, the board trusts the director. A hallucinated figure breaks that chain at the top, where it is most visible and hardest to repair.
The cost is not the error itself. Errors get corrected. The cost is the next ten minutes of the meeting, when everyone wonders what else in the pack is wrong. After that, a director will check every AI output by hand, and a tool that has to be checked line by line has stopped saving time.
**AI does not lose a banker's trust slowly. It loses it in one meeting.**
A number without a source is a liability, not an answer
The fix is not a smarter model. It is a rule: no figure without its source. Every ratio, rate and peer number should link back to the filing, report or regulator release it came from, with the page where possible.
Regulators are moving the same way. MAS's AI risk management guidelines, issued on 7 October 2026, keep financial institutions accountable for third-party AI they use. OJK's AI governance guide for Indonesian banks, launched on 29 April 2025, asks banks to run AI safely and with effective risk management. Neither accepts "the AI said so" as an answer.
**If you cannot click through to where a number came from, it is not ready for a board.**
Dates matter as much as sources
A sourced number can still embarrass you if it is stale. Interest rates, deposit costs and peer results move every quarter. A correct figure from last quarter, presented as current, is a wrong figure.
So the second rule is a date on everything that moves: "as of Q2 2026", "per the 30 September filing". It sounds pedantic until the day someone across the table quotes the newer number.
**In a boardroom, the right number from the wrong quarter is still the wrong number.**
Fix the boring failures before adding features
Hallucinations get the headlines, but bankers lose patience with smaller things first. A reply that lands in the wrong conversation. The same message sent twice. A summary that ignores the instruction given five minutes earlier.
None of these is dramatic. Each one tells a careful person that the tool is not careful. For an assistant that lives in a chat app, reliability is not a back-end metric. It is the product a director experiences every day.
**Senior users forgive a missing feature. They do not forgive a sloppy one.**
Answer the question, skip the lecture
The last source of embarrassment is the assistant itself. Bankers ask precise questions and expect precise answers. An assistant that adds unrequested disclaimers, moral caveats or a paragraph on "consulting a professional" makes a senior person feel managed.
Flag a real risk when there is one. Otherwise, answer what was asked, show the source, and stop. That is the standard we hold Rapi™ to, and the one we would hold any AI analyst to. For how to test it, see how to test an AI analyst.
**The goal is not an AI that sounds careful. It is one that makes you look careful.**
Sources
- Monetary Authority of Singapore: MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial Institutions (7 October 2026)
- Monetary Authority of Singapore: Consultation Paper on Guidelines on Artificial Intelligence Risk Management
- Otoritas Jasa Keuangan: Peluncuran Tata Kelola Kecerdasan Artifisial Perbankan Indonesia (SP 67/GKPB/OJK/IV/2025)
- Otoritas Jasa Keuangan: Tata Kelola Kecerdasan Artifisial Perbankan Indonesia
Common questions
- What is an AI hallucination in banking?
- An AI hallucination is a figure, fact or citation an AI tool presents confidently but that is not supported by any source. In banking, the damaging version is a ratio, rate or peer number that ends up in a board pack or committee discussion and turns out to be wrong or out of date.
- How can banks reduce the risk of AI hallucinations?
- Set two rules for any AI output used in decision papers: every figure must link to its source document, ideally to the page, and every figure that changes over time must carry a date. Then test the tool on real questions before rollout, and keep a person accountable for what is forwarded.
- Does MAS regulate banks' use of third-party AI tools?
- Yes. MAS's Guidelines on AI Risk Management, issued on 7 October 2026, apply to all financial institutions and all forms of AI. Institutions remain accountable for third-party AI they use and should obtain sufficient assurance from providers. The expectations start applying from 7 October 2027.
