Table of Contents
- Key takeaways
- What does a chatbot do in customer service?
- Why should a chatbot start with a narrow set of jobs?
- Which questions should a chatbot answer, and which should go to a person?
- How to use a chatbot for customer service, job by job
- Job 1: Answer the questions your team gets every day
- Job 2: Cover the hours nobody is online
- Job 3: Triage conversations and hand off the hard ones
- Job 4: Point order and account questions to the right place
- Where should you start? A decision matrix
- How do you measure a customer service chatbot?
- Worked example: one month of chatbot numbers
- Common mistakes when using a chatbot for customer service
- Where SiteGPT fits
- Frequently asked questions
- Getting started
- How do you use a chatbot for customer service?
- What should a customer service chatbot answer first?
- How long does it take to roll out a customer service chatbot?
- Handoff and after-hours
- Can a chatbot hand a conversation to a person?
- What should a chatbot do when nobody is online?
- Orders and accounts
- Can a chatbot tell a customer where their order is?
- Should a chatbot change account details?
- Measurement
- How do you know if a customer service chatbot is working?
- What is a good chatbot resolution rate?
- Sources

- Start with one job. Train the chatbot on your help content and the 20 questions your inbox gets most, then test it on those 20 before launch.
- Add after-hours cover next. A team that works 40 hours a week leaves 128 of the 168 hours in a week (76%) without a person online.
- Write handoff rules before you widen the scope. In a 2026 Gartner survey of 3,566 customers, 87% said access to a human agent is essential.
- Keep order and account questions honest. A chatbot trained on content can explain the returns policy, but it cannot see one customer's order unless it is connected to the order system.
- Track five numbers every week: resolution rate, handoff rate, fallback rate, answer rating, and time to a human reply.
Key takeaways
Job | What the chatbot does | What it should leave to a person | Number to watch |
1. Answer repeat questions | Answers from your help center, policies, and product pages | Questions your content does not cover | Resolution rate |
2. Cover after-hours | Answers when the team is offline and sends an away message on handoff | Anything that cannot wait for the next working day | Handoffs answered by the next working day |
3. Triage and hand off | Spots requests for a person, urgency, and frustration, then passes the transcript | Refunds, complaints, billing disputes | Handoff rate and time to a human reply |
4. Order and account questions | Explains policies and links the tracking or account page | Any answer that needs a lookup the chatbot cannot do | Fallback rate on order questions |

What does a chatbot do in customer service?
Why should a chatbot start with a narrow set of jobs?
Which questions should a chatbot answer, and which should go to a person?
Question type | Example | Who answers | Why |
Policy or how-to | "What is your return window?" | Chatbot | The answer is on a page and is the same for everyone |
Product detail | "Does the annual plan include phone support?" | Chatbot | The answer is on the pricing or product page |
One customer's order | "Where is order 48213?" | Chatbot links the tracking page, or a person looks it up | The answer needs a lookup in the order system |
Account change | "Change the email on my account." | Person, or a link to the account settings page | The change needs the customer's identity confirmed |
Refund or billing dispute | "I was charged twice." | Person | A wrong answer costs money |
Complaint | "This is the third time it broke." | Person | The customer needs a person, not an article |

How to use a chatbot for customer service, job by job
Job 1: Answer the questions your team gets every day
- Train the chatbot on the pages that answer those 20 questions: help articles, policies, the pricing page, and product pages.
- Write any missing answers as short pages or text snippets. A chatbot cannot answer what nobody wrote down.
- Ask the chatbot all 20 questions in your own words. Mark each answer correct, partly correct, or wrong.
- Fix the content behind every wrong answer, then run the 20 questions again.
Job 2: Cover the hours nobody is online
- Set the team's working hours and time zone in the chatbot.
- Choose what happens outside those hours: the AI answers, or the team answers later.
- Write an away message that says when a person will reply. Do not promise a reply time the team cannot keep.
- Check the overnight handoffs every morning and reply to each one.
Job 3: Triage conversations and hand off the hard ones
- Turn on handoff so a visitor can ask for a person in one click or one message.
- Write escalation rules in plain language. Name the topics that always go to a person, such as refunds, billing disputes, legal threats, and complaints.
- Make sure the transcript travels with the handoff, so the customer does not repeat the question.
- Decide who on the team picks up handed-off chats, and during which hours.
Job 4: Point order and account questions to the right place
- Train the chatbot on every policy that order questions touch: shipping times, returns, exchanges, and cancellations.
- For live order status, have the chatbot link the tracking page or the account page.
- For account changes and anything it cannot look up, hand off and ask the customer for the order number first.
- If your tool offers an order lookup, test it on 10 real orders before you switch it on for customers.
Where should you start? A decision matrix
If you need | Start with | Why |
Fewer "same question again" tickets | Job 1: answer repeat questions | The top 20 questions usually carry the most volume |
Answers for customers in other time zones | Job 1, then job 2 in the same week | After-hours cover only works once the answers are right |
Faster replies on hard cases | Job 3: triage and hand off | The team sees the transcript and skips the first questions |
Fewer "where is my order" chats | Job 4, starting with a link to the tracking page | A link resolves the question without an order lookup |
A chatbot in a regulated service, such as healthcare | Job 1 with strict handoff rules | Keep the bot on public information and route anything personal to a person |
How do you measure a customer service chatbot?
Metric | How to calculate it | What a good trend looks like |
Resolution rate | Conversations closed without a person, divided by all conversations | Rising week over week |
Handoff rate | Conversations handed to a person, divided by all conversations | Stable, with handoffs on the topics you chose |
Fallback rate | Conversations where the bot said it did not know, divided by all conversations | Falling as you add missing content |
Answer rating | Helpful ratings, divided by all rated answers | Rising, with low ratings read one by one |
Time to a human reply | Time from handoff to the first reply from a person | Inside the time your away message promises |
Worked example: one month of chatbot numbers

- Resolved: 780 of 1,200 conversations, a resolution rate of 65%.
- Handed off: 264 conversations, a handoff rate of 22%. Over 22 working days, that is 12 handoffs a day for the team.
- Fallback: 156 conversations, a fallback rate of 13%. These are the questions the content does not answer yet.
Common mistakes when using a chatbot for customer service
- Hiding the route to a person. Customers who cannot reach a person leave the chat and email instead, so the work comes back.
- Letting the chatbot guess order status. A confident wrong answer about an order costs more than a link to the tracking page.
- Never updating the content. Prices, policies, and shipping times change. Refresh the trained content whenever the source pages change.
- Reading only the dashboard. Low ratings and fallback answers explain themselves only in the transcripts.
- Promising instant replies at night. The away message should name the real reply time.





























