A customer asks your chatbot, “Can I change the delivery address after payment?” The bot searches its approved business information but finds no clear answer. This is the moment that separates a useful assistant from a frustrating one. A responsible chatbot should not guess, repeat a generic sentence or make the customer start again. It should recognise the exception, be clear about its limits and move the enquiry towards the right person.
That does not mean every uncertain question needs an immediate live takeover. It means the business must design the fallback before the chatbot goes live. The goal is continuity: the customer knows what will happen next, and the team receives enough context to help.
1. Recognise when an answer is not reliable
A chatbot needs defined boundaries. It should be able to identify questions outside its approved knowledge, conflicting information, missing details and requests that require a decision by a member of staff. The trigger might be an unsupported topic, a low-confidence response, a repeated misunderstanding or a sensitive complaint. The exact technical rule depends on the system, but the business rule should be simple: do not present an unverified answer as fact.
Before launch, list the situations that always require review. These might include refunds, unusual pricing requests, account access, complaints, urgent service failures or anything involving information the bot is not authorised to disclose. Test those cases deliberately instead of testing only easy questions.
2. Say what the chatbot can and cannot do
A good fallback is direct and useful. For the delivery-address example, the chatbot might say:
“I don’t have a confirmed answer for changes after payment, so I don’t want to give you the wrong information. I can pass this to our customer team. Please confirm your order number and the best way to reach you. A team member will review the request and reply during business hours.”
This illustrative wording does not promise a response time that the business has not set. It also avoids pretending that the bot has completed a change. If the customer has already supplied the order number, the chatbot should not ask for it again. A handoff should reduce repetition, not create another form of friction.
3. Transfer context, not just a name
The human team needs the conversation that led to the escalation: the customer’s original question, relevant answers, order or booking reference where appropriate, and the reason the bot stopped. A short summary can help, but it should be checked against the transcript before a staff member relies on it. The customer should not have to explain the entire situation again.
Engage24’s AI Switchboard & Chatbots service describes CRM integration with conversation transcripts, contact information and next steps, as well as escalation of serious prospects. Those stated capabilities are useful reference points when discussing what context a proposed implementation should preserve. The exact routing, permissions and response process still need to be agreed for each business.
4. Define ownership and exception queues
“Handoff to a person” is not a complete process unless someone owns the queue. Decide who receives each type of exception, during which hours, what happens outside those hours and how urgent cases are identified. Give staff a way to correct the answer, record the outcome and flag knowledge that needs updating.
NIST’s voluntary AI Risk Management Framework emphasises managing AI risk across design, development, use and evaluation. Its Playbook also calls for clear roles and responsibilities for human-AI oversight, and for monitoring systems in production because conditions can change. For a small business, that can translate into a practical routine: name an owner, review escalations, sample conversations and pause or change the bot when recurring failures appear.
5. Measure the exceptions and improve the boundary
Do not judge the chatbot only by how many conversations it completes. Track unanswered questions, repeat misunderstandings, handoff reasons, time to human response and whether customers had to repeat themselves. Review a sample of both successful answers and escalations. A spike in one question may indicate missing business information, a confusing flow or a process that should never have been automated.
The best fallback is not an admission of failure. It is a controlled route to expertise. When a chatbot can answer from reliable business knowledge, it should do so. When it cannot, it should be honest, preserve context and make human responsibility visible.
If you are assessing where a customer-facing chatbot could fit, you can request a tailored quote from Engage24 or review the AI Switchboard & Chatbots service. Start by mapping the exceptions your team already handles; those cases will often tell you more about a safe design than the easy questions do.
Sources
- NIST AI Risk Management Framework — voluntary framework overview and lifecycle risk-management purpose.
- NIST AI RMF Playbook: Govern — documented roles, responsibilities and human-AI oversight.
- NIST AI RMF Playbook: Measure — production monitoring, unexpected data and human review.
- Engage24 AI Switchboard & Chatbots — verified service description and stated integration, escalation and monitoring claims.

