Chatbot

Most bank chatbots launched between 2017 and 2022 followed the same arc: high-profile launch, modest adoption, quiet retirement. The failure was not the concept but the implementation. Keyword-matching systems trained on FAQs cannot handle the ambiguity of real customer queries. Large language models have changed what is technically possible, but regulated financial services adds constraints that the technology alone does not solve.

What is a chatbot?

A chatbot is a software application that conducts text-based or voice-based conversation with a user. In financial services, chatbots sit on banking apps, insurer portals, and internal service desks, handling queries that range from balance checks to mortgage application status updates. The term covers a wide spectrum: from rule-based decision trees that follow scripted paths to conversational AI systems that understand natural language and generate contextual responses.

The distinction between a chatbot and a virtual agent matters for regulated firms. A chatbot that provides account information is delivering a service. A chatbot that recommends a savings product or suggests a customer increase their insurance cover may be crossing into regulated advice. The regulatory boundary is not about the technology. It is about the outcome. If the customer acts on the chatbot's output as if it were a recommendation, the firm bears the same responsibility as if a human adviser had spoken.

Adoption rates in UK banking vary widely. Some institutions report 40 to 60 per cent of digital queries handled by chatbot. Others struggle to reach 15 per cent. The difference is rarely the AI model. It is the depth of integration with core systems. A chatbot that can tell you your balance but cannot initiate a payment, update an address, or raise a complaint is a novelty. One that can execute transactions within the conversation is a channel.

The landscape

The FCA's Consumer Duty has sharpened the requirements for customer-facing chatbots. Firms must ensure that automated channels deliver outcomes at least as good as human channels. This includes accessibility for customers with disabilities, appropriate handling of vulnerable customers, and clear escalation paths when the chatbot cannot resolve the issue. A chatbot that loops a distressed customer through the same FAQ without escalation is a Consumer Duty failure.

Generative AI has reset customer expectations. Before ChatGPT, users accepted that chatbots required specific phrasing and could only answer predefined questions. Now they expect natural conversation, contextual memory, and accurate responses to open-ended queries. Banks that deployed rule-based chatbots three years ago face a choice: upgrade the technology or accept declining usage as customers compare the experience unfavourably.

Internal chatbots for staff are growing faster than customer-facing ones. Compliance teams, operations staff, and relationship managers use internal chatbots to query policy documents, search knowledge bases, and draft responses. The regulatory bar is lower because the user is a trained professional, and the output is reviewed before reaching a customer. Several UK banks report higher ROI from internal chatbots than from customer-facing deployments. The EU AI Act's transparency requirements apply to customer-facing chatbots, requiring firms to disclose when a customer is interacting with an AI system.

How AI changes this

Retrieval-augmented generation (RAG) is the pattern that makes generative chatbots viable in regulated environments. Rather than generating answers from the model's training data (which may be outdated or wrong), the system retrieves relevant content from an approved knowledge base and uses the language model to formulate a natural response. The answer is grounded in verified content. The model provides the conversational fluency. This approach addresses the hallucination risk that makes unconstrained LLMs unsuitable for financial services.

Multi-turn conversation with memory is now achievable. A customer can say "I want to check a payment I made last Tuesday to my landlord," and the system parses the date, identifies the likely transaction, and presents it for confirmation. Follow-up questions ("Was it the one for 850 pounds?") maintain context. This conversational depth was impossible with rule-based systems and is what drives meaningful adoption.

Sentiment detection within the conversation enables dynamic escalation. If the chatbot detects frustration, confusion, or distress, it can transfer the customer to a human agent with full conversation context. The handover is seamless rather than forcing the customer to repeat themselves. This directly supports Consumer Duty obligations around vulnerable customer identification.

What to know before you start

Define the boundary between service and advice before writing a single line of code. Map every possible customer query against your regulatory permissions and determine which queries the chatbot may answer, which it must escalate, and which it must decline. This mapping is a compliance exercise, not a technology exercise, and it determines the chatbot's scope. Your data governance framework should define which customer data the chatbot can access and under what conditions.

Test with real customer language, not internal terminology. Customers say "my card isn't working" when they mean a dozen different things: declined transaction, lost card, expired card, frozen account, wrong PIN. Your training data must reflect the ambiguity of real queries, not the precision of your internal taxonomy. Conversation logs from your existing channels are the best source.

Plan for graceful failure. The chatbot will encounter queries it cannot handle. What happens next determines whether the customer's experience is good or terrible. A clear message ("I can't help with that, but let me connect you to someone who can") with a warm transfer that carries context is infinitely better than "I didn't understand that. Please try again." Build the fallback path before building the happy path.

Integrate with knowledge management from day one. A chatbot is only as good as the content it draws from. If your knowledge base is scattered across SharePoint sites, PDFs, and tribal knowledge, the chatbot will reflect that fragmentation. Consolidate and structure your knowledge before expecting the AI to make sense of it.