Fintechs are automating more of their customer support with AI, and the usual measure of success is how much contact no longer needs a person. In regulated financial services, a short customer message can carry vulnerability, fraud or consent issues that an automated answer is not equipped to handle.
Anastasia Ioseliani previously worked at a UK regulated fintech, where she progressed into customer experience management and worked directly with AI-supported customer operations. In this opinion piece she draws on that experience, without reference to any confidential company or customer information, to argue that the more important question is when AI should stop and escalate. The views are her own.
The customer’s message looked simple enough. They could not make a payment. For an automated support system, this is exactly the kind of query that seems easy to handle. Identify the topic, locate the relevant information, generate the response and move on.
But anyone who has worked in regulated customer support knows that a short message can contain much more than a short question. Why can’t the customer make the payment? Have they lost their job? Are they dealing with a bereavement? Is someone else controlling their finances? Are they confused about what they owe, or are they telling us that they simply cannot afford it?
The technical answer may be easy. The correct response may not be. That difference is where I think much of the conversation around AI in customer service is still missing the point.
Companies are understandably focused on what AI can answer. In regulated industries, we should be paying just as much attention to what it should refuse to answer.
The obsession with automation
AI has obvious value in customer support. A large percentage of customer enquiries are repetitive. Customers want to know where to find something, why a transaction is pending, how a process works or what they need to do next. These are areas where automation can work extremely well. It can reduce waiting times, remove repetitive work from support teams and give customers access to information almost instantly.
Having worked with AI-supported customer service processes in regulated fintech, I am not sceptical about the technology itself. Quite the opposite. I have seen how useful it can be.
What concerns me is the assumption that the natural endpoint of good automation is more automation. It is easy to start measuring success by the percentage of conversations that no longer require a person. But there are situations where avoiding human involvement should not be the goal. Sometimes escalation is the correct outcome.
Customers rarely speak in compliance language
One of the hardest parts of customer support is that customers do not describe their circumstances using the categories companies use internally. A customer rarely writes: “I am experiencing financial vulnerability and require additional support.”
They say: “I can’t pay this week.” They say: “My partner normally deals with all of this.” They say: “I’ve been off work for a while.” They say: “I don’t understand any of these charges anymore.”
A human support agent may immediately recognise that the conversation needs more care. An automated system may simply identify the apparent question and continue.
This is particularly important in financial services because vulnerability is rarely contained in one obvious keyword. Context matters. Tone matters. The history of the conversation matters. Sometimes what looks like a routine payment question is no longer a routine payment question once you understand what sits behind it.
AI can be trained to recognise certain indicators, but recognition alone is not enough. The system also needs rules for what happens next. In some cases, the correct next step should not be a better automated answer. It should be: Stop. Escalate this conversation.
Fraud is where ‘helpful’ can become dangerous
Fraud-related conversations are a good example of why an AI system cannot simply be trained to provide the fullest possible explanation. Customers naturally want to understand what is happening. Why was a payment stopped? Why is additional verification required? Why has an account or transaction been reviewed? What exactly caused the system to flag something?
Those questions are completely reasonable from the customer’s perspective. But in fraud prevention, more information is not always better. There are situations where explaining exactly how a fraud control works, what triggered a review or which internal indicators were detected could make the system less effective.
An automated assistant that is heavily optimised around being transparent and helpful may not understand that distinction unless the boundaries are deliberately built into it. This is one of the areas where refusal matters.
The AI should not try to fill in the gaps. It should not speculate about the reason for a fraud review. It should not reveal internal detection logic. It should not confirm assumptions simply because the customer phrases them confidently. Sometimes the correct response is deliberately limited.
That can feel uncomfortable in customer service, because we are used to thinking that a better explanation always creates a better experience. In regulated environments, that is not always true. A very detailed answer can be operationally worse than a careful one.
Data, consent and the temptation to answer because the information exists
There is another area where AI needs very clear boundaries: customer data. Support teams often have access to large amounts of information. Identity details, transaction histories, account activity, previous conversations and sometimes information provided by third parties can all form part of a customer record.
The fact that information exists inside a system does not automatically mean it should be used, repeated or shared in every conversation. Who is asking? Has their identity been properly verified? Are they asking about their own information? Are they acting on behalf of somebody else? Do they have authority to do so? Has the customer actually consented to their information being shared?
These questions are easy to overlook when an AI model can retrieve information instantly. Imagine a family member, partner or representative contacting a company and asking for an update on somebody else’s account. The system may have the answer. That does not mean it should provide it.
The same issue appears when a customer casually mentions another person during a conversation. An automated system should not treat every piece of information available to it as fair game simply because it is technically accessible. This is why data protection cannot be reduced to a disclaimer at the bottom of a chatbot. The permission to access information and the permission to disclose it are two different things.
In practice, good automation needs to understand not only what information it knows, but under what circumstances it is allowed to use that information. And when there is uncertainty around identity, consent or authority, the safest response may again be to stop. Not because the AI lacks the answer. Because it should not be the one giving it.
Information is not the same as financial advice
There is another boundary that matters in FCA-regulated financial services: the difference between providing information and giving a customer a recommendation. A customer may ask: “Should I make this payment now or wait?” “Which option is better for me?” “What should I do with this balance?”
From a customer-service perspective, those questions can sound completely ordinary. But depending on the product, the firm’s regulatory permissions and the context of the conversation, there may be an important difference between explaining what options exist and telling the customer what they personally should do.
That distinction becomes even more important with AI. AI is naturally good at producing recommendations. Give it a few facts and it will often try to identify the ‘best’ option, explain why and present the answer confidently. In a regulated environment, that instinct can create risk.
An automated system should be able to explain factual information clearly: what a payment option means, when something is due, what the consequences of a particular process are, or where the customer can find further information. But it should not automatically turn that information into personalised financial advice where the firm, product or interaction does not permit it.
The difference can be surprisingly small in language. “There are three available options” is information. “Based on what you’ve told me, you should choose the second option” can be something very different. That is exactly the kind of line an AI system may cross without realising that it has crossed it.
In FCA-regulated businesses, customer communication is not just about whether an answer sounds helpful. Firms also need to consider whether communications are fair, clear and not misleading, whether vulnerable customers are being treated appropriately, and whether the interaction stays within the regulatory permissions and responsibilities of the business. This is another situation where the safest AI may be the one that knows when to stop giving an answer.
We should design AI to fail safely
A lot of AI product design focuses on reducing failure. That makes sense, but in regulated environments the definition of failure needs to be more careful. Refusing to answer is not necessarily failure. Escalating is not necessarily failure. Admitting uncertainty is not necessarily failure. The real failure may be continuing confidently when the system does not have enough information or should not be making the decision.
AI can still play a significant role. It can summarise long conversations. It can surface relevant information for agents. It can identify repeated customer issues. It can categorise straightforward enquiries. It can help teams understand where customers are getting stuck. It can reduce enormous amounts of repetitive work. None of that requires the system to become the final decision-maker in every interaction.
The strongest use of AI in regulated customer support may not be replacing judgement. It may be helping humans know where judgement matters most.
The post When Should AI Refuse to Answer in Regulated Fintech Support? appeared first on The Fintech Times.
