What should I do when a customer asks a question the chatbot cannot answer?

For most small businesses, the practical goal is not to add more software. It is to make the next step clear, reliable, and easy for both staff and customers.

Document the facts the system may use, the actions it is allowed to take, and the situations that must reach an employee. Keep prices, hours, policies, and service descriptions in an approved source rather than relying on improvised answers. Test both routine requests and edge cases, and make the handoff include enough context that customers do not have to start over.

For this decision, consider safe fallback language, clarifying questions, escalation, logging unanswered questions, and updating the knowledge base after review.

For example, an assistant may answer approved questions about hours or booking steps, but route complaints, unusual requests, sensitive information, and questions requiring professional judgment to a person.

1. Describe the current state. Write down who starts the process, what information is collected, which tools are involved, and how the work ends.
2. Choose a specific improvement. Examples include fewer missed follow-ups, fewer scheduling conflicts, more reliable records, or clearer reporting. Set a baseline before changing the process.
3. Check fit and constraints. Confirm that the existing tools, plan levels, account permissions, APIs, and platform rules support the intended workflow. Keep a working tool when a targeted integration or process change will solve the problem.
4. Pilot with realistic cases. Include ordinary use, mistakes, missing information, cancellations, duplicate submissions, and failures. Ask the people doing the work to test it.
5. Assign an owner. Decide who reviews results, handles exceptions, updates source information, and responds when something breaks.

NIST's Artificial Intelligence Risk Management Framework and Generative AI Profile are useful references for treating AI risk management as an ongoing process of design, evaluation, monitoring, and improvement. OWASP's GenAI security guidance also highlights risks such as prompt injection, sensitive-information disclosure, and giving AI systems excessive authority. These risks should be assessed against the actual tools and permissions in use.

Automation and AI can produce incomplete or incorrect responses. Do not let an assistant invent prices, promise refunds, make professional judgments, or perform consequential actions without appropriate controls. NIST's AI Risk Management Framework and its Generative AI Profile recommend a risk-management approach across design, use, evaluation, and monitoring.

Other avoidable mistakes include automating a process nobody has agreed on, collecting more customer data than necessary, failing to define who is responsible for the next action, and measuring activity instead of the outcome the business actually wants.

• Define approved information; test common and difficult questions; set clear escalation triggers; limit access to customer data; log failures; review real conversations and update the process.
• Confirm access, privacy, security, and any applicable legal or platform requirements.
• Record a baseline and agree on what success means.
• Test the workflow end to end and document a manual fallback.
• Review results after launch and improve the process based on real exceptions.

Can a chatbot transfer a conversation to a real employee? (Article 35), How can I create a reliable FAQ knowledge base for my business? (Article 33), What are the risks of connecting AI to customer conversations? (Article 45)

If you want to improve this process without replacing tools that already work, EASY can help assess the workflow and scope a practical solution through AI and Business Automation. Use the Talk to us button to describe your current setup, the problem you want to solve, and the tools your team uses, then request a quote. Feasibility and scope will depend on your workflow, software access, and integration options.

---