How to implement an AI chatbot with mandatory handoff of complex inquiries to a human

AI support chatbot answers user questions and passes nonstandard inquiries to an employee. Without this handoff, automation becomes a barrier: people receive repetitive answers, do not understand how to ask for help, and stop trying to resolve the problem.

The British organization Citizens Advice called for preserving access to a live employee, especially for matters involving medicine, social assistance, utilities, and interactions with government services. For a service owner, this is not only a matter of interface convenience. The escalation process determines whether a user receives help when a standard answer is not suitable.

Which inquiries can be entrusted to a chatbot

An AI chatbot works best with simple informational requests when the answer is described in advance in a knowledge base and does not change the user’s legal, financial, or medical status. This group includes information about opening hours, required documents, appointment procedures, payment methods, application status, and general service rules.

• The chatbot provides informational content when the answer does not depend on an individual decision by a specialist.
• The system helps users find the relevant section, application form, or instructions.
• Automation collects initial information before transferring the inquiry to support.
• The bot displays process status when it retrieves it from a verified information system.

The boundary is not between the user’s “simple” and “complex” text, but between an informational prompt and a decision that affects a person. A question about office hours can be automated. A decision about a charge, benefit, medical recommendation, social payment, or service termination requires the involvement of an authorized employee under the specific organization’s rules.

Topics where the bot must not make the final decision

Before launch, compile a list of restricted scenarios. Include it in the project documentation rather than leaving it as an informal agreement between the product team and support service. The system identifies such topics by keywords, operation type, risk level, and conversation context, then offers to transfer the inquiry to an employee.

• Medicine: symptoms, deterioration of health, interpretation of examination results, and questions about the urgency of seeking care.
• Social assistance: allocation of payments, confirmation of eligibility for support, complaints about denial, and situations involving a person’s vulnerable circumstances.
• Bills and payments: disputed charges, arrears, service disconnection, refunds, and changes to personal terms.
• Government services: appeals against decisions, refusal to accept documents, missed deadlines, and inquiries requiring an official response.
• Safety and crisis situations: threats to life, violence, loss of access to a critically important service, and other cases where delay increases the risk.

The chatbot can explain the general procedure, but it must not present an assumption as a specialist’s decision. If a user asks for an assessment of a specific entitlement, assignment of assistance, or confirmation of a medical conclusion, the interface must clearly state the limits of the automated response and offer an appropriate contact channel.

How to make the handoff to an employee noticeable

The “Contact an employee” button should be located in the conversation window, not hidden in the help section. A clear handoff option should be available at any time, especially after an answer that did not solve the problem. A command such as “agent” or “human” should work without a long menu.

For important categories, the handoff can be shown immediately after the first message. For example: “If your question concerns a bill, medical assistance, or a social service, you can contact an employee right away.” This wording reduces the risk that a person will spend time on an unsuitable process.

What happens if a user has already received the same answer several times? The conversation should end the loop and offer an employee rather than continue generating new wording. Repeated expression of the same intent, short messages such as “I don’t understand,” a negative rating of the answer, and an explicit request for help serve as escalation signals.

Which signals trigger the handoff of an inquiry

Escalation rules are best described through both the content of the request and the user’s behavior. An intent classifier alone is insufficient: a person may not use the expected keyword, describe the problem in everyday language, or change the subject after several unsuccessful attempts.

The model’s confidence threshold must not be the sole criterion. Even a confidently worded answer may be unsuitable if the topic concerns rights, money, health, or an individual decision. In such scenarios, a business rule is more important than the model’s probabilistic assessment.

What to transfer to the agent along with the inquiry

The employee should see the problem before joining the conversation. The system transfers the correspondence history, inquiry topic, the user’s answers to clarifying questions, selected documents or forms, information about previous inquiries, and the reason for escalation. The user should not have to retell the situation simply because the automated process has ended.

Data collection is limited to the purpose of the inquiry. The chatbot must not request information that is unnecessary for answering or transferring the request. Access to the history is granted only to employees who need it to process the inquiry, while the service operator determines data retention periods and deletion procedures in accordance with applicable personal data requirements.

A good message for the agent looks like this: “The user disputes the charge for May, the personal account number has been confirmed, the bot showed general instructions, and the user requested a review.” Such a summary saves time but does not replace the original correspondence if the details affect the decision.

How to account for the digital divide

Online chat is not suitable for all users. Citizens Advice specifically noted that some people do not have stable internet access, a modern smartphone, or a computer, while others find digital services difficult to use. If an organization leaves only a chatbot available, it excludes people who need help no less than anyone else.

• Keep a telephone channel if the service is connected with important life decisions.
• Provide contact through an office, mail, or another accessible method appropriate to the service profile.
• Display alternative channels next to the chatbot, not only in the legal section of the website.
• Do not require users to register again when they move from the chat to another channel.
• Describe the operating conditions of each channel in plain language, without assuming technical skills.

An alternative channel is needed not only when there is no internet access at all. It is important during outages, when the connection is slow, for accessibility, and in situations where a person finds it difficult to formulate a request in writing.

Which metrics to check after launch

The number of conversations and share of automated responses do not show whether the system helped. The support team tracks how many inquiries ended without a resolution, how many users requested an employee, how often the chatbot repeated answers, and how many people returned with the same problem.

• share of inquiries transferred to an employee;
• share of conversations ending without a clear result;
• number of repeat inquiries on the same topic;
• frequency of repeated answers and loops without progress;
• share of users who ended the conversation after an error or refusal;
• time until an employee joins;
• assessment of the result after the inquiry is transferred;
• number of cases in which the alternative channel was unavailable.

Citizens Advice cited data according to which almost half of respondents considered chatbots useless, while 14 percent stopped trying to get help after an unsuccessful conversation. These figures relate to the survey cited by the organization, not to a universal result for all chatbots. For the project team, the conclusion is clear: an unfinished conversation should be counted as a loss of outcome, not as successful automation.

How to test the support service

Testing typical questions shows only how the informational layer works. Before launch, the team adds nonstandard inquiries: incomplete descriptions, contradictory data, repeated messages, emotional wording, errors in an application number, requests to appeal a decision, and transitions between several topics.

1. Create an inquiry map and identify categories where an employee’s decision is required.
2. Describe the rules for ending the conversation and the conditions for mandatory transfer.
3. Configure a button, command, and other ways to contact a human.
4. Transfer the history, collected data, and reason for escalation to the agent.
5. Test the telephone and other alternative channels using the same scenarios.
6. Launch a limited pilot and manually review unfinished conversations.
7. Change the knowledge base, routing, and interface based on the causes of errors, not only on user ratings.

Also test the situation in which an employee is unavailable. The chatbot should honestly state the operating hours, offer to leave an inquiry or choose another channel, and preserve the information entered. Promising immediate help without a real queue creates a new source of distrust.

Frequently asked questions

Do we need to abandon chatbots entirely?

No. Automation is suitable for informational questions and preparing an inquiry. The limitation concerns not the chatbot itself, but its use instead of an employee where an individual decision is required or the user cannot complete the standard process.

When should the handoff to a human be available?

The handoff should be available from the start of the conversation and after any sign of failure: misunderstanding, a repeated answer, a negative reaction, a request for an agent, or an inquiry on a sensitive topic. The user should not have to prove their right to help by going through the same process several times.

Can a brief summary be sent to the agent instead of the entire correspondence?

A summary speeds up the work but does not replace the history if the details affect the decision. The employee should see the original messages, collected data, and the reason the bot stopped automated processing.

How can we tell whether the implementation was successful?

Success is determined not by the maximum share of automation but by the outcome for the user. The system should answer simple questions quickly, end unsuccessful processes in a timely manner, transfer complex inquiries to an employee, and preserve an accessible alternative channel.