A customer asks their bank's virtual assistant about early mortgage payoff penalties. The system needs to pull the current policy, interpret it correctly, and answer within seconds. Get it wrong, and the bank faces regulatory exposure. Refuse too often, and the assistant becomes useless.
That tension sits at the center of a new study on production banking AI, carried out by a team from the research department at Backbase, and led by Denys Katerenchuk, Head of AI Research at Backbase.
Published at the Annual Meeting of the Association for Computational Linguistics (ACL) in July 2026, the study describes a large language model already deployed across more than 40 financial institutions. The findings challenge some of the industry's default assumptions about how AI governance works in regulated environments.
The core finding: the difference between a safe, production-ready banking AI and an unsafe or unusable one came down to calibrated refusal, achieved by deliberately engineering the training data rather than by reaching for a bigger or more advanced model.
That's worth unpacking for any bank still deciding on how to approach governance for customer-facing AI. For the broader framework this study fits into, see our complete guide to AI governance in banking.
The problem: how banking AI agents fail customers
Most large language models fail banks in one of two directions:
They hallucinate: Faced with a question they can't fully answer, they generate a plausible-sounding response anyway. In banking, an incorrect answer about a fee, a rate, or a policy carries real compliance and financial risk.
They over-refuse: Trained to be cautious, they decline to answer even when the information is available, frustrating customers and defeating the purpose of the assistant.
The study measured this directly. An untuned base model answered almost everything, refusing only 4.3% of the time, an unsafe pattern given how often those answers were unsupported. A leading general-purpose model, by contrast, refused 20.2% of the time, erring so far toward caution that it became unhelpful on questions it could have answered correctly.
Finding one: refusal is a governance decision
The study's central technical contribution is a method for deliberately teaching a model when to say "I don't know."
The researchers built a training set where 22% of examples paired a customer question with context that looked relevant but didn't actually contain the answer. The correct response, in every one of those cases, was an explicit refusal.
The result: a model that abstained 12% of the time, a rate that sits between the base model's dangerous overconfidence and the general-purpose model's excessive caution. Push the share of unanswerable examples in the training mix past roughly 26%, and the resulting model starts refusing questions it could actually answer. There's a real, measurable sweet spot in how that training data gets built.
For banks, this reframes refusal. Rather than a limitation to route around, refusal becomes a governance mechanism that needs tuning as deliberately as any policy rule, and something worth asking any AI vendor to show data on.
Finding two: sequencing the training data mattered more than the model itself
The most counterintuitive result in the study: when the researchers combined all their training data into a single, unsequenced pool instead of feeding it in stages, performance collapsed. Answer quality dropped by more than 40% on the study's scoring scale, and the model started over-refusing on almost half of all queries, most of which it could have answered.
Critically, data volume played no role here. Both approaches trained on the same 98,648 samples. The only variable was order: teaching the model general financial language first, then specific, calibrated refusal behavior second, versus mixing everything together at once. The staged, curriculum training version produced the best overall balance of answer quality and safe refusal behavior. A separate version, trained on external data alone, actually scored slightly higher on two narrower measures, precision and citation quality. But it did so at the cost of almost never refusing anything, a pattern the researchers flag as unsafe on its own.
The lesson: sequence and curate the data you have, whatever its size. This cuts against a common instinct in banking technology circles, where bigger models and larger datasets are often treated as inherently better. The study suggests process matters more than raw scale.
Finding three: real users prefer an honest "I don't know" over a confident guess
Evaluation benchmarks only go so far, so the study checked real user feedback in a real setting: seven months of customer interactions at a large U.S. credit union serving retail banking customers.
The results moved in the same direction on both fronts. Resolution rates improved by 7.1 percentage points, a statistically significant gain that meant fewer customers had to escalate to a human. Satisfaction improved too, from 59.5% to 62.9%, though that particular increase wasn't statistically significant on its own and would need a longer study.
Put those together and the implication is notable. This is a model that refuses to answer nearly three times as often as its untuned base, yet customers resolved more of their issues and, if anything, came away slightly happier. In a real banking setting, a calibrated "I don't know" beats a confident hallucination: an honest refusal keeps the customer's trust and routes them to a correct answer, while a fabricated one merely feels resolved until it isn't.
Many banks measure AI success through response quality scores or sentiment analysis on individual interactions. This finding suggests resolution rate, the blunt measure of whether the customer's problem actually got solved, is the more reliable metric to watch. It's also the one the data can back up with statistical confidence.
Finding four: grounded AI doesn't have to be expensive or slow
The purpose-built model ran 3 to 5 times faster than a leading general-purpose model, at roughly 20 to 50 times lower cost per query. Even after compressing the model down to roughly a third of its original size for efficient deployment, it retained over 99% of its original citation quality. And the results held beyond the researchers' own data: on FinanceBench, a public benchmark of SEC filing questions, the model outperformed both its untuned base and the general-purpose comparison.
That combination matters because it removes a common excuse for slow AI adoption in banking: the idea that accuracy and safety require expensive, heavyweight infrastructure. A smaller, well-trained, purpose-built system beat a much larger general-purpose one on both quality and cost.
What this means for AI governance in banking
Pull these findings together and a clear set of priorities emerges for any bank building or buying AI for customer-facing or operational use:
- Treat refusal as a governed behavior. Ask how a system's abstention rate was tuned, and what happens when it's wrong in either direction.
- Don't equate more data or bigger models with better outcomes. Sequencing and curating the same training data outperformed dumping it all in at once.
- Put resolution rate at the center of how you measure success. A model that answers fewer questions but answers more of them correctly may outperform one that sounds better but resolves less.
- Question the assumption that safety costs more. The most accurate, best-governed system in this study was also the fastest and cheapest to run.
The pattern holds beyond any single bank, vendor, or use case, and points to what works when AI meets the accuracy and compliance demands of financial services: evidence in place of assumption. The full study is available on arXiv for anyone who wants the underlying methodology and data.
FAQs
What does AI governance in banking mean for AI models?
It means treating behaviors like refusal, citation, and abstention as deliberately governed and measured, instead of leaving them as side effects of a model's default settings. The study found that calibrating exactly when a model says "I don't know" was as important as its answer accuracy.
Does more training data always improve AI model performance?
Not according to this study. Two approaches used the exact same amount of training data. The one that sequenced it in stages significantly outperformed the one that combined it all at once. With volume held constant, order and curation made the difference.
What actually drives customer satisfaction with banking AI assistants?
Getting issues resolved. In seven months of real credit union data, a model that refused more often but hallucinated less resolved significantly more queries and showed a measurable satisfaction gain, suggesting customers prefer an honest "I don't know" over a confident wrong answer.
Is accurate, well-governed banking AI more expensive to run?
Not necessarily. The purpose-built model in this study was both more accurate and 20 to 50 times cheaper per query than a general-purpose frontier model.
