The AI chat will become the most frequent form of interaction that many customers have with their bank, insurer or wealth manager – and financial institutions are investing accordingly. But are these chats being built to serve customers, or to deflect them? In the third of our six-part series, we explore what trust means in the AI chat context and set out six principles for creating AI chats that customers are drawn to, rather than quickly learning to avoid.
A UK insurer was testing a new AI chat ahead of launch. The team was confident the experience was ready. Customer testing said otherwise. The most common feedback was not about wrong answers. Customers told them they did not trust the responses they were getting.
When the insurer engaged us to look closer, the AI’s reasoning was sound. The answers were defensible. The problem was that customers could not see what the AI was doing or how it had arrived at any given response. The chat was a black box.
We redesigned the experience to do two things. Show the customer in real time what the AI was doing. Explain the reasoning behind each answer in plain English.
Same AI. Same underlying logic. Different interface. At the next round of testing, customer trust scores rose sharply. That experience captures everything this article is about.
Most AI chat interfaces in UK financial services look identical. A box. A blinking cursor. A polite greeting. Underneath, very different briefs.
Every UK bank, insurer and wealth manager is building one of these right now. The teams designing them are making choices that will define the customer experience for the next decade.
Some are built to genuinely help the customer. Some are built to keep the customer away from a human. The customer can usually tell within two exchanges.
The interface often reveals the brief. Key signals include whether and how the chat shows its workings, what it lets the customer do, and when it hands over to a person. These look like design choices but they send a strong message about the institutional intentions behind the provision of the AI chat.
Even well-intentioned institutions can mistakenly send the wrong signals to customers, if they get their interface wrong. Six principles separate a chat that customers come back to, from one they learn to avoid:
A customer asking about a missing payment does not want a confident answer that turns out to be wrong. They want to see what the AI is actually doing. Is it searching their account, or guessing from training data? Most current chats do not tell them.
Good chat interfaces show this. Bad ones hide it.
The principle is the same one that applies to a good human adviser. You trust the answer more when you can see how it was reached. The customer does not need a technical explanation. They need to see that something real is happening behind the response.
The insurer experiment showed why explainability sits alongside transparency, not below it. Once customers could see what the AI was doing, the next thing they wanted to understand was the reasoning behind any given answer.
Take a typical interaction. A customer asks why their insurance renewal premium has gone up by £85. The AI gives them the new figure with a one-line reason: “increased due to changes in your risk profile.”
The customer’s next question, every time, is “which bit, and by how much?”
A good chat can walk through it. How much of the increase came from a single claim. How much from a postcode-level risk reassessment. How much from market conditions. How much from optional cover the customer added at last renewal.
The customer might still feel the increase is too high. They will at least understand what they are being charged for.
A bad chat gives the headline and stops. The customer is left with a number they has not been explained and that, for example, they cannot defend to their partner. In financial services, that is where trust starts to wear thin.
This is the difference between a chat that feels like a conversation and one that feels like a wall.
Customers need to be able to interrupt the AI. Stop a long response. Cancel an action. Undo something that was just done.
This sounds obvious. It is missing from most chats currently in production.
The reason is that interruption is harder to engineer. It is easier to let the AI finish. Easier to commit the action immediately. Easier for the bank, worse for the customer.
If a customer cannot stop the AI mid-flow, they will hesitate before using it for anything that matters. And that defeats the point.
Some customers want a personalized experience. Some find it intrusive. Most are somewhere in between, and their preference shifts depending on what they are trying to do.
The chat should reflect this, for example, by offering a clear opt-in to AI features. The ability to disable personalization without losing access to the service. Honest, plain-language information about how data is being used and what the customer can change.
A common failure mode is the AI that gives a perfect answer and then leaves the customer to do all the work themselves.
“You can update your address through online banking.” Fine. But the customer is in the chat. Why are they being sent somewhere else?
How those actions happen also matters. Customers are comfortable giving instructions conversationally. They are less comfortable confirming consequential actions through chat alone.
A customer can tell the chat to send £200 to a friend without thinking twice. What they want before pressing send is a visual confirmation: a clear panel inside the chat showing the recipient, the amount, the date, the reference. A button to confirm. A button to cancel. The same pattern they already trust from their mobile banking app.
This is where multimodal design matters. The conversational layer takes the instruction. A visual layer, designed to feel like the existing mobile experience, confirms the action before it happens.
Good chat experiences combine both. They are a place where things actually happen, with the same level of trust the customer has in the rest of the bank’s digital experience.
The single most important question for any AI chat in financial services concerns when it should hand over the conversation to a skilled human professional.
Some interactions should never be fully automated. A vulnerable customer in distress. A complex claim. A regulated advice conversation. A complaint that has already escalated.
The chat needs to recognize these moments and route to a human, ideally without the customer having to ask. Confidence thresholds should trigger a human review. Sensitive topics should default to a human agent. The handover itself should be smooth: the human picks up where the AI left off, with full context, without making the customer repeat themselves.
Most banks are still treating the chat as the destination. The right design treats it as a layer, with a clear and well-rehearsed exit to a human.
For most customers, the AI chat will become the most frequent interaction they have with their bank, insurer or wealth manager. More frequent than the app, the branch and the call center.
If they do not trust your AI chat, they do not trust you.
The technology to build the right kind of AI chat exists today. The constraint is the brief.
Which brings us back to the question that runs through this whole series. Are chats being built to serve customers, or to deflect them? The interface will tell you – and your customers – the answer.
To discuss how Capco can help your organization design and implement trusted AI chats within a framework of AI Orchestrated Experiences, speak to our experts.
