Chatbots are now common across major banks, but customer trust remains uneven. Fewer run one their customers actually trust. Deloitte surveyed 2,027 customers of US banks in January 2025 and found that 37% of respondents had never interacted with banking chatbots, despite years of investment and near-universal deployment.
That gap reflects a confidence problem. A chatbot customers don't trust gets bypassed, escalated, or abandoned. A chatbot customers don't trust gets bypassed, escalated, or abandoned, and each of those outcomes costs the bank money and goodwill.
The cost of getting conversational AI wrong
A poorly trusted chatbot doesn't fail quietly. The cost shows up in four places.
Customer cost
Deloitte surveyed 2,027 customers of US banks in January 2025. The survey found that 74% preferred human agents over chatbots for routine banking queries.
That preference creates a direct customer cost. People remain willing to use chatbots for quick answers and simple requests, but many still choose human support when confidence matters. Among respondents who preferred chatbots, 57% wanted better accuracy. Deloitte also found that chatbot users valued quick responses at 78% and simple-query handling at 69%.
Customers therefore pay with time and effort when a chatbot cannot provide accurate, useful help. They move to a phone call, branch visit, or human chat. The bank then pays for another service channel. The practical measure is the share of routine requests customers complete through automation without human escalation or repeat contact.
Operational cost
A 2025 case involving a large Australian bank shows what can happen when staffing plans rely on projected chatbot performance. The bank introduced an AI voice bot and announced plans to remove 45 contact-center roles. The bank had said the voice bot would reduce call volumes by 2,000 calls each week. After the rollout, reported call volumes increased. The bank reversed the staffing decision after managers arranged overtime and moved team leaders onto customer calls.
The lesson applies beyond voice bots. Automation does not remove operational work unless customers complete their requests successfully. When demand shifts to employees, the cost appears through overtime, redeployment, longer queues, and repeated handling.
Business cost
The business risk of getting conversational AI wrong sits between projected efficiency and measured outcomes.
A projected reduction in call volume is not the same as a completed customer request. Banks need evidence that customers received accurate answers, completed their tasks, and avoided repeat contact before treating automation as a saving.
The Australian bank's case shows why that distinction matters. A staffing plan built around an expected reduction had to change after reported call volumes increased. The bank then revised its staffing approach.
For any conversational AI deployment, track:
- Completed customer tasks
- Human escalation
- Repeat contact
- Cost per completed outcome
- Customer satisfaction
- Channel switching
These measures show whether automation creates genuine capacity or moves unresolved work elsewhere.
Risk and reputation cost
Broader research on generative AI in banking, separate from chatbot-specific use, points to a trust ceiling that shapes how far banks can push automation. Deloitte reported that 72% of banking customers were concerned about sharing financial information with gen AI tools, that 83% would feel anxious about an AI agent taking action on their finances without approval, and that 66% said the risk of AI errors could outweigh the benefits. A chatbot that gives wrong information, oversteps its authority, or acts without clear consent can increase frustration and reduce trust.
What customers expect
Customers expect a chatbot to finish the job: an accurate answer on the first try, a clear path to a person when the bot can't help, and no need to repeat information already given. The blog on why customers expect conversational banking breaks down that bar in more detail. The consistent thread across the research is: customers judge a chatbot on whether it solved their problem, not on how the conversation felt getting there.
Why chatbots lose trust
Most trust failures come down to answers that are wrong or outdated, missing visibility into account history, or scripts that ignore what the customer already said. According to Deloitte's research, 57% of respondents who preferred chatbots said better accuracy was the improvement they most wanted to see.
Much of this traces back to what the chatbot can see. A conversational system built on top of fragmented banking systems can only answer the questions it has data for, and defaults to a generic script for everything else. AI chatbots in banking covers how these systems are built and what separates real account context from a system guessing at an answer.
Banking chatbot accuracy and wrong answers covers the technical causes behind incorrect responses, from stale data to model hallucination.
When chatbots should hand off to people
Escalating to a human protects trust when the handoff carries full context forward. Trust breaks when customers have to repeat everything they already explained to a second channel.
A chatbot should hand off, with context intact, when:
- The request involves a dispute, a complaint, or a request for an exception.
- The customer expresses frustration or repeats the same request.
- The action is consequential, such as moving money or closing an account, and needs an approval step.
- The bot's confidence in its own answer is low, or the topic touches sensitive financial guidance.
- The customer directly asks for a person.
A customer forced to re-explain their problem to a second channel has effectively lived through two bad experiences instead of one.
What chatbot containment measures
Containment gets reported to leadership because it's easy to measure. It also measures the wrong thing. A chatbot can contain a conversation, meaning the customer never reaches a human, without solving the customer's problem at all. Containment without resolution is where trust erodes fastest, because the bank records success while the customer experiences a dead end.
A bank optimizing for containment alone will keep reporting improvement while trust keeps falling. Banking chatbot use cases covers how to instrument these metrics at the point of deployment.
Responsible deployment
Trust holds up when customers know what they're talking to, understand its limits, and stay in control of anything consequential. That starts with disclosing when a customer is speaking with an AI system rather than a person, and being clear about where the system's authority ends.
The bank sets that boundary. Some actions, like a balance check or a statement request, can execute without friction. Others, like moving money or closing a product, should require explicit customer confirmation before anything happens. That distinction determines whether customers keep using the chatbot or route around it. See the Backbase Conversational Banking demo to explore a governed conversational experience.
FAQs
Do customers trust banking chatbots?
Trust is mixed and depends on the task. Deloitte surveyed 2,027 US banking customers and reported that 37% of respondents had never interacted with banking chatbots, while 74% preferred human agents over chatbots for routine queries.
Why do banking chatbots frustrate customers?
Accuracy is the top source of frustration. Deloitte reported that 57% of respondents who preferred chatbots wanted better accuracy, and rigid scripts, limited account visibility, and handoffs that force customers to repeat information - plus long wait times when escalation is needed - add to the frustration.
When should a banking chatbot hand off to a human?
A banking chatbot should hand off to a human when the request involves a dispute or complaint, the customer is frustrated or repeats a request, the action needs approval, the bot's confidence is low, or the customer asks for a person. Passing the conversation history at handoff reduces the need for customers to repeat themselves.
How should banks disclose AI use to customers?
Banks should tell customers when they are interacting with an AI system and clarify which actions it can complete on its own versus which need explicit approval. In a broader 2026 survey on generative AI in banking, Deloitte reported that 83% of respondents would feel anxious about an AI agent taking action on their finances without approval.
Is a high containment rate the same as good chatbot performance?
A high containment rate only shows that a conversation stayed automated. Task completion measures whether the request was actually resolved, so banks should also track first-contact resolution, repeat contact rates, CSAT, and cost per completed outcome.
