Conversational Artificial Intelligence (AI) is becoming an integral part of financial services, as well as for other industries. As organizations are developing tools for employees and customers, one design challenge consistently emerges: deciding when, and how, an AI system should refuse a request. Santander teams are currently working on different tools to overcome this challenge while ensuring finding the right balance.
Human interactions need to be observed from the design stage of any AI systems, especially when they are related to a conversational platform. Therefore, the AI system should find the right balance to ensure acceptance and refusal of human requests. That engineering challenge walks through a thin line. If AI refuses too often, it may become frustrating and difficult to use. In the other hand, if it refuses too little, it can be providing unreliable information or create harm.
Banco Santander is working on this dilemma, developing several measures that the system must comply with to be well balanced. Luis Ibáñez, Head of Responsible AI at Santander, underlines that “as we worked extensively on testing our conversational AI solutions, we found that understanding how customers naturally interact with the system across specific use cases is essential to avoiding unnecessary refusals. By studying the language they use, the context they provide and the expectations they bring, we can better calibrate the system, reducing over-refusal while maintaining the safeguards needed for trustworthy AI”.
The sweet spot
Systems that reject reasonable requests too frequently undermine customer trust and reduce the value of the service. On the other hand, public examples have shown how overly permissive systems can generate inaccurate information, make commitments they were never authorised to make or produce responses that damage an organisation's reputation.
In every case, responsibility remains with the organisation deploying the technology. For financial institutions, where conversations may influence decisions involving money, eligibility or advice, the stakes are even higher.
Understanding why refusals matter requires looking beyond technology. Human conversation follows shared social expectations. Research in linguistics and sociology helps explain why. Cooperative conversations depend not only on accuracy but also on relevance, clarity and sufficient explanation. Refusals are naturally sensitive moments that people typically soften by explaining their reasoning, suggesting alternatives and preserving the other person's dignity. Conversational AI should follow the same principles. A refusal can remain firm while also being clear, proportionate and genuinely helpful.
Another challenge is less visible but equally important. As conversational systems become more fluent, users increasingly respond to them as social actors rather than software. Research shows that some people, particularly those who are isolated or vulnerable, can develop strong emotional attachments to AI assistants. The technology does not need genuine understanding for those relationships to have real effects.
This places an additional responsibility on organisations deploying conversational AI. Systems should be approachable and easy to use without encouraging users to mistake them for people. Transparency is therefore essential, not only to meet regulatory requirements but also to preserve trust.
The regulatory landscape reflects this direction. The European Union's AI Act requires users to be informed when interacting with AI in many contexts, with transparency obligations for conversational systems applying from August 2026. Similar approaches are emerging in other jurisdictions, reinforcing common expectations around accountability, disclosure and the protection of vulnerable users.
Designing trustworthy conversational AI therefore requires more than deciding what a system can or cannot do. It requires designing refusals that are clear, proportionate and helpful while ensuring the system remains transparent about its capabilities and limitations. Trust is built not by avoiding difficult conversations, but by handling them honestly, consistently and with respect.
