Gains:
- Being able to distinguish which tasks can be safely accelerated by chatbot and artificial intelligence-supported response in customer services and where delegation to humans is necessary.
- Ability to use AI to draft answers, summaries and knowledge base searches, managing the risks of misinformation, unauthorized transactions and disclosure of secrets
- Ability to set rules to stop automation and redirect to authorized employee in cases requiring complaints, sensitive requests and financial advice
Customer services are the place where the bank contacts the customer most frequently: balance inquiry, password reset, campaign information, complaint, transaction objection, product application. Most of these requests are repetitive, and AI-powered chatbot and response assistants can handle them in seconds. But customer service in banking is not an innocent information desk: incorrect information leads to financial loss, an unauthorized transaction leads to fraud, and revealed information leads to a violation of customer secrets. So the basic principle is: the chatbot speeds up the routine; complaint, sensitive advice and authorized action are delegated to the human. Knowing where automation will stop is just as important as installing it.
In this unit, we will see what tasks the chatbot does safely, where handover to humans is necessary, and how to manage the risks of misinformation and secret leakage.
What can he do safely and what can't he do?
Request type
chatbot
Why
Balance/transaction query (authenticated)
does
routine, verifiable
FAQ: opening hours, product conditions
does
Static, low risk
Password reset guidance
Does (with safe streaming)
standard process
Complaint and dissatisfaction
transfers to human
Empathy and solution authority
Investment/credit advice
transfers to human
Compliance and liability
Authorized transaction in the account (high amount transfer, etc.)
Additional verification + human
Fraud risk
Unclear/complex situation
transfers to human
Risk of misinformation
The rule is simple: chatbot speeds up low-risk, standard and verifiable work; Jobs that involve emotion, advice, high authority or uncertainty go to people.
Tip: The most important feature of a good chatbot is its ability to say "I don't know, I'm connecting you to an official." The chatbot that tries to answer every question is the most dangerous.
Three big risks and their management
- False information (hallucination). The model may make up a non-existent campaign or the wrong interest rate. Solution: connect the chatbot only to the approved knowledge base and ensure that it does not give an answer that does not have a source.
- Revealing customer secrets. The chatbot can provide account information to an unauthenticated person or move a customer's data to another session. Solution: authentication and session isolation before every sensitive operation.
- Unauthorized transaction. The chatbot completes transactions such as money transfers on its own. Solution: requiring additional verification and human approval for high-risk transactions.
Response assistant: artificial intelligence alongside the employee
While the chatbot speaks directly to the customer, there is also a response assistant that helps the employee: it puts the response draft, relevant procedure and summary in front of the customer representative. The rule is the same here — the assistant produces a draft, the agent checks it and sends it. The difference is that one filters each answer; Therefore, the response assistant can be used more flexibly than the chatbot that speaks directly to the customer.
Four copyable templates
1) FAQ answer draft (based on knowledge base):
Your role: assistant who DRAFT the response to the customer representative. Rely solely on the approved knowledge base text I gave you; Any terms, rates or promotions that are not there are FAKE. If there is no answer in the knowledge base, say "this issue should be forwarded to the authorized person". Knowledge base: [related product text] Customer question: [question]
2) Transfer and summary for the complaint:
Your role: assistant who classifies incoming messages. Is this message a complaint, dissatisfaction, or request for financial advice? If so, tag "TRANSFER TO HUMAN" and provide a neutral 2-sentence summary of the situation for the representative. Don't come up with solutions or advice yourself. Message: [customer message]
3) Empathetic but limited response outline:
Your role: assistant who proposes drafts to the agent.Customer complains about a delayed transaction. Draft a response that is non-accusatory, empathetic, but DOES NOT promise a definitive solution/date. The representative will review the transaction; Just set up a polite hold and redirect text.
4) Security check (sensitive request):
Consider the following customer request: does it require authentication or involve an authorized transaction (transfer, limit change, information update)? If so, list what additional verification step is required and issue an "automatic conclusion NOT APPROPRIATE" warning. Request: [request]
Weak prompt / Strong prompt
Weak prompt:
Always answer customer questions precisely and quickly, if necessary, make a reasonable estimate so that the customer does not have to wait.
The “predict” instruction invites hallucination, does not connect to the source, and ignores handover points.
Powerful prompt:
Your role: response assistant based only on approved knowledge base. Do not make up anything that is not in the knowledge base. Otherwise, forward it to the authorized person. Mark complaints, recommendations and authorized transaction requests as "TRANSFER TO HUMAN". Do not provide account information without verification of identity. The representative will approve the draft.
Strong will prohibit fabrication, define handoff points, and prevent secret leakage.
three mini cases
Case 1 — Routine speed gain. A bank's chatbot answers 500,000 FAQ-type questions per month (working hours, card delivery status, campaign condition) from its approved knowledge base. Agents are left with only the complex 20%; waiting time decreases from 8 minutes to 2 minutes. Questions that do not have a source are automatically directed to people.
Case 2 — Hallucination is prevented. A customer asks, "Do you have a 0-interest loan campaign?" In a wrongly designed system, the model could make up "yes, this month has 0 interest". The chatbot, which is tightly connected to the knowledge base, says "There is no such campaign in the knowledge base, I am connecting you to the authority." Misinformation and possible legal problems are prevented.
Case 3 — Unauthorized transaction is stopped. A customer tells the chatbot, "send 50,000 TL from my account to another account." The chatbot classifies this as a high-risk authorized transaction and does not auto-conclude it; leads to additional verification and authorized employee approval flow. This step makes it harder to exploit a compromised account.
Designing the turnover point: smooth transition
Handover to the human is part of the proper functioning of the chatbot, not its “failure.” But a poorly designed handoff infuriates the customer: telling the same information over again, long queues, the line saying "I'm transferring you to the relevant unit" and then disconnecting. A good cycle carries three things:
- Context transfer: The information collected by the chatbot (who the customer is, what he asked, what steps were tried) is briefly passed to the representative; The customer does not repeat himself.
- Expectation management: The customer is clearly told why he is being conveyed and how long he will wait. Uncertainty is the main source of dissatisfaction.
- Irreversible transaction halt: In a situation where handover is required, the chatbot does not attempt to complete the transaction; He leaves it unfinished and transfers it safely.
Write down the rollover thresholds in advance: what keywords (e.g. "complaint", "lawsuit", "I was scammed", "investment advice"), what transaction types and after how many unsuccessful attempts will the automatic rollover be triggered? Leaving these rules vague will cause the chatbot to persist in the wrong place.
Tip: One of the best metrics to measure the quality of a chatbot is the “time-to-time turnover rate.” Too late turnover tires the customer; Not transferring it at all is the most dangerous. Transfer is not a defect, it is a safety valve.
Common mistakes
- Forcing answers to every question. Setting up the chatbot so that it cannot say "I don't know"; the main source of hallucination.
- Skipping turnover points. Not directing complaints and advice to people.
- Providing information without authentication. Exposing sensitive account information to an unauthenticated session.
- Automating authorized process. Leaving transactions such as transfers and limit changes to the chatbot alone.
- Not keeping the knowledge base up to date. Producing wrong answers with old rate/campaign information.
Attention: Incorrect information given by a chatbot may be legally binding on the bank. “The system said so” is not a defense; Therefore, the chatbot should only provide approved, up-to-date and sourced information, and hand over the rest to the human.
In summary
Chatbot and response assistants expedite routine, standard, and verifiable requests; Complaint, sensitive advice, competent action and uncertainty are delegated to the human. The three main risks—misinformation, secret leakage, and unauthorized transaction—are managed by an approved knowledge base, authentication, and human confirmation, respectively. The most valuable chatbot feature is that it knows its limits and can say "I am connecting you to the authority" in a timely manner. In one sentence: AI speeds up speech; The authorized person finalizes the sensitive decision and transaction.
Application task
Write five typical customer requests (e.g. balance inquiry, promotion question, complaint, transfer request, credit advice request). For each, determine whether it is “chatbot does” or “hands off to human” and why. Then produce an FAQ response with template 1 and a complaint handover summary with template 2. Test the "redirect rather than make up" behavior by asking a question that is not in the chatbot's knowledge base.
checklist
- [ ] I have correctly classified the request as low risk / human transfer.
- [ ] I have only connected the chatbot to an approved, up-to-date knowledge base.
- [ ] I have blocked sharing sensitive information without authentication.
- [ ] I delegated complaints, recommendations and authorized action to humans.
- [ ] I put additional verification and human approval on high-risk transactions.
- [ ] I took into account that the responsibility for the answers lies with the bank.