Automation & AIModule 3: Business workflowsLesson 12 of 15
Course progress73%
20 min lesson · Updated August 2026
How do chatbots support customers?
A support chatbot can guide navigation, answer approved low-risk questions and collect context, but it needs clear identity, source boundaries, privacy controls, accessibility, logging and a fast human escalation path.
What you will learn
By the end, you will understand:
Choose between rule-based, retrieval and generative support designs
Define safe answer and action boundaries
Create accessible escalation and service-quality evaluation
Visual explainer
See the idea clearly.
✓
A customer question enters a transparent support assistant; simple navigation follows rules, knowledge questions use approved sources, sensitive or uncertain issues move to a human with conversation context.
Different bot designs solve different jobs
Design
Use
Rule-based
Known menu choices, status paths and deterministic FAQs with predictable behavior.
Search/retrieval
Find relevant approved help articles and let the customer inspect sources.
Generative RAG
Synthesize an answer from approved knowledge with citations and no-answer behavior.
Agentic support
Take scoped actions such as checking status, only with authentication, permission and confirmation.
Tell people what they are using
Do not pretend the bot is a human. State that it is automated, what it can help with, how data is used and how to reach a person. Avoid human names/avatars that materially deceive users.
A bot should not claim emotion, professional qualification or authority it does not have.
Collect less conversation data
Tell users not to enter passwords, full payment-card details or unnecessary sensitive information. Redact/minimize logs, define retention and limit staff/vendor access.
Authenticate before showing account-specific information. Conversation possession alone does not prove identity.
Escalate by risk and failure
01
Understand request
02
Check source/permission
03
Answer approved low-risk question
04
Detect uncertainty/sensitive topic/failed action
05
Offer human route
06
Pass concise context with consent
07
Human accepts ownership
08
Close and learn
Escalation triggers
Safety or emergency
Medical/legal/financial advice
Threat, abuse or vulnerability
Complaint or cancellation
Identity/account security
Payment/refund beyond limit
Repeated misunderstanding
No approved source
Tool failure
User requests a person
Accessibility barrier
Language unsupported
Accessibility and conversation control
Use semantic controls, keyboard access, visible focus, screen-reader labels and sufficient contrast. Do not trap focus or force speech. Let users pause, review, copy and correct information.
Messages should be concise, and errors should explain recovery. A person must not be forced through many failed loops before human help.
Measure resolution, not deflection alone
Signal
Question
Task success
Was the user’s problem actually resolved?
Escalation quality
Did the human receive accurate context and take ownership?
Answer quality
Were sources current and citations relevant?
Safety/privacy
Were sensitive inputs, leaks or unauthorized actions detected?
Experience
Could users understand, control and exit the interaction?
Real-world example
Example: delivery-status assistant
Example
The bot explains it is automated, authenticates the customer, checks one order-status API and never changes delivery addresses. Missing status or a complaint creates a human ticket with the customer’s permission. It does not invent a delivery time.
Try this
Write a bot boundary card
List ten supported questions, ten escalation triggers, prohibited data, allowed tools, authentication method, source owner, retention, human hours and the exact no-answer message.
Common questions
Questions beginners ask.
What is a chatbot?
A conversational interface that may use fixed rules, search, language models or tools to help users.
Should a bot say it is automated?
Yes. Users should understand the nature and limits of the interaction.
Can a chatbot answer from company documents?
Yes with governed retrieval, permissions, source maintenance and no-answer behavior; errors remain possible.
When should a bot hand over to a person?
For sensitive, high-impact, unsupported, failed, repeatedly misunderstood requests or whenever the user asks.
Can a bot show order details immediately?
Only after appropriate identity/authentication and authorization checks.
Should chatbot logs be kept forever?
No. Define necessary purpose, access, redaction and retention, considering applicable law and contracts.
What is containment or deflection?
A conversation completed without human support; it is not success if the user’s problem remains unresolved.
Can a chatbot replace emergency support?
No. Emergency/safety contexts need clear appropriate human or emergency routes.
