Website chat after hours: what an AI agent answers, and what it hands to a person
An AI agent in your website chat can answer customers at night. It answers from what your business has checked, and passes anyone ready to book to a person on your team.
On this page
- What should an AI agent in my website chat actually do?
- Why did a heli-ski operator want a chat that answers after hours?
- How is an AI agent different from a chatbot on my website?
- What did the agent need to know first?
- What should an AI agent hand to a person?
- How do you test an AI agent before real customers use it?
- What did real visitors show us that testing did not?
- What should you check before adding website chat with an AI agent?
- Questions people ask about AI agents on websites
- It answers only from what your business wrote down and checked.
- It shows options as buttons, because many visitors tap instead of type.
- Every booking goes to a person, with the whole chat attached.
- Test it with fake visitors first. Real visitors will still find new things.
Website chat with an AI agent can answer your customers at night, when nobody is in the office. It needs two things to work: answers your business has checked, and a quick way to reach a real person. We built one for a heli-ski operator, and here is what we learned.
What should an AI agent in my website chat actually do?
It should answer the questions your team answers every week, at any hour, and pass every booking to a person.
That comes down to four jobs:
- Answer from what you know. Your products, your rules, your way of saying things. No guessing.
- Show what fits. If a visitor asks about options, show them right in the chat.
- Speak the visitor’s language. If you sell abroad, questions come in more than one language.
- Hand over to a person. When the visitor is ready, or the question is too hard, someone on your team takes over.
Most chatbots on websites get the last job wrong. A chat that cannot reach a person is a dead end.
Why did a heli-ski operator want a chat that answers after hours?
Because its guests look at trips at all hours. A question that waits until morning is often a booking made somewhere else.
The operator sells multi-day trips to guests from all over the world, with a small office team. So questions often arrive while the office is asleep. Our report for the operator showed that most people the agent talked to came outside office hours.
During the day, the team answered the same questions again and again. Which trip suits my level? What is included? Is there space that week? The agent was built to take those repeat questions. The team kept the bookings and the decisions. Read the full case study.
How is an AI agent different from a chatbot on my website?
An AI agent understands the question the visitor typed and answers in plain words. An old-style chatbot only follows fixed menus.
What matters to your visitor: can it deal with the question they asked, and does it get them to a person when needed?
It only answers what it was set up for and goes in circles on anything else. The visitor gives up or calls in the morning.
- Fixed menus and stock replies
- No way to reach a person, or only an email address
- Your team starts every enquiry from scratch
It answers in plain words from what the business checked, and hands the visitor to a person when it should.
- Answers the question the visitor actually asked
- Shows the trips or products that fit, inside the chat
- A person gets every booking, with the whole chat
What did the agent need to know first?
It needed the operator’s own knowledge, written down and checked by the operator, before any visitor could talk to it.
Together with the operator, we wrote down the trips, the lodges, safety and how to prepare, plus answers to the most common questions. The operator read it all and made changes. Then the team gave feedback three more times.
This is the slow part, and it should be. An AI agent is like a new employee. It gets better because you keep adjusting it to what visitors really ask.
The chat also had to feel like the operator’s own website. Visitors can open and compare the trips, the lodges and the weather right inside the chat. Many visitors never typed a question at all. They just tapped the buttons.
What should an AI agent hand to a person?
Every booking, every request to talk to a person, and every question it cannot answer well.
When a visitor wants to book or asks for a person, the agent opens a short contact form. The team gets an email with the request and the whole chat, and the visitor gets one too.
| What the visitor wants | What the agent does | What a person on the team does |
|---|---|---|
| Which trip suits my level, what is included, how to prepare | Answers from what the operator checked | No need, unless the visitor asks for them |
| To see the trips, the lodges or the weather | Opens them inside the chat | – |
| An answer in German, French or Spanish | Answers in that language | – |
| Dates and open places | Explains how booking works and passes the visitor on | Takes it from there |
| To book | Opens the form and sends the team the chat | Confirms every booking |
| To talk to a person | Passes them on straight away | Picks up knowing what was said |
| Something the agent cannot answer | Says it does not know and passes it on | Answers it |
| Details about the operator's helicopters | Politely says no | – |
How the heli-ski agent splits the work.
Visitors asked most about dates and open places. Showing open dates inside the chat is the next thing to build. If you want this for your own business, see how we build booking agents.
How do you test an AI agent before real customers use it?
You let AI pretend to be different visitors, score every chat, and fix what you find before launch.
Pick a test visitor below to see how they talked to the agent and what the testing found.
Pick a test visitor. None of them were real.
Before a real guest typed a word, AI played four kinds of visitor, and every chat was scored on six things.
Asks whether this trip is even for someone like them.
Pushes hardest on Tone and Helpful.
What this visitor checksChecks that the answers stay plain and friendly, and that a person is offered when the agent is unsure.
Pushes on the detail the brochure skips.
Pushes hardest on Right facts and Odd questions.
What this visitor checksChecks that every detail is right, and what the agent does with a question it cannot answer.
Books for people who ski at different levels.
Pushes hardest on Handover and Helpful.
What testing foundThe contact form opened exactly when it should, and the group went to a person.
Keeps asking for things a rival would want to know.
Pushes hardest on Safety and Right facts.
What testing foundFound three things, all fixed before launch: a wrong price, details about the helicopters it should not share, and a handover that did not start.
The chart is the average of ten test chats, a month before launch. Out of 5.
We also added simple safety checks, so nobody can talk the agent into saying something it should not. Personal details are removed before chats are saved. And the chat only works on the operator’s own website.
What did real visitors show us that testing did not?
That launch day is the start of the work. Real visitors find things no test will.
On launch day, our own checks found a bug that would have stopped the booking button from working. We fixed it the same afternoon, before any visitor saw it.
Within days, real visitors showed us two more: a greeting that appeared twice, and a form that could be sent twice. We fixed both.
Adding it to the website was the easy part: one line of code on the operator’s WordPress site.
The agent went live in the last weeks of the 2025/26 season, so we saw a short launch window, not a full season. Nobody used the thumbs up or thumbs down buttons, so we read nothing into them.
A month after the season ended, the agent was still passing new enquiries to the team.
What should you check before adding website chat with an AI agent?
Check that the same questions keep coming in, that someone owns the answers, and that a person is ready to take over.
Before you put an AI agent on your websiteTick off what is already true for your business./8 done
If most of these are true, an agent will probably save your team real time. If your answers are not written down yet, start there. Not sure where AI fits in your business at all? Read where to start with AI in your business, or book a free call.
Questions people ask about AI agents on websites
Is website chat worth it for my business?
Yes, if customers contact you outside office hours and your team answers the same questions over and over. If you sell to people in other time zones, whoever answers first usually gets the booking. If you get few questions and they are all different, a clear contact page may be enough.
Can an AI agent take bookings on its own?
It can collect everything a booking needs and send it to your team, with the whole chat attached. We build it so a person confirms every booking. The agent prepares, and your team decides.
What stops an AI agent from making things up?
It only answers from what your business has written down and checked. When a question goes beyond that, it says so and passes the visitor to a person. We test this before launch with fake visitors who ask odd and tricky questions, and we keep checking after.
Can I make my own AI agent for free?
Yes, free tools exist that let you try a simple chatbot on your website. Making it good takes more work: writing down and checking what it should know, testing it with tricky visitors, deciding when it hands over to a person, and adjusting it once real visitors arrive.
Does an AI agent replace the person who answers enquiries?
No. It takes the repeat questions and the late-night ones, so your team can spend its time on the chats that turn into bookings. Anyone who asks for a person gets one.