An AI agent for customer support that answers from your own branch list
Branch addresses, opening hours and the nearest location are most of what customers ask. An agent that answers them from your tables and passes on the rest.
Part of the guide AI employees
An AI agent for customer support earns its place on the questions that never change shape: where is your nearest branch, what time do you close today, which location has parking, is the one on Main Street open on public holidays. Put your branches in a table, give the agent that table as a tool, let it check its own answers before they go out, and send what it cannot answer to a person. In xMap Flows that is Chat (or a WhatsApp message), the AI agent with My data, Check the work, Reply, and Ask a person.
This guide is for service and operations teams of multi-site businesses: gyms, car rental desks, pharmacies, service centres, banks with branches. It is less about clever conversation and more about getting dull facts right every time.
Start with the table, not the AI
Most wrong answers from support bots come from missing facts, not weak AI. Before building anything, put one row per location in a CSV and upload it to My data: name, address, phone, weekday hours, weekend hours, holiday hours, parking (yes or no), services offered. My data reads the latitude and longitude columns if you have them, or you can add a position to each row from its address when you bring a list in.
Longer things that do not fit a row, such as your returns policy or membership terms, go in a document the agent searches with Look things up in. It answers from the passages it finds and does not make the rest up.
The flow, block by block
- Chat starts the flow when someone writes on your chat page, shared by link or placed on your website with the code in the flow's menu.
- AI agent, with instructions such as You help customers of our gym chain find a branch and its opening hours. Answer only from the branch table and the membership terms. If you do not know, say a colleague will follow up. Link My data (the branch table) to its Tools dot, and Nearby places if people ask what is around a branch.
- Check the work, with rules like every answer that names a branch gives its phone number, no hours that are not in the data and the ready-made check against made-up numbers. An answer that fails goes back to the agent with the reason.
- Reply, which sends the checked answer back into the chat.
To find the nearest branch the agent needs to know where the customer is. Write the customer's place into the tool's point as {{agent}}; when someone says near the night market in Chatuchak, the place name is looked up for you and the agent can compare distances.
Protection on a public page
Anyone with the link can write anything. The Chat block's Protection settings, under More options, handle the common problems: a topic list so questions about unrelated things get a polite no, refusal of messages that try to change its rules or reveal its setup, and removal of personal data from answers, except the contact details you publish. When the chat page is switched on, the last two start on. The agent applies these together with any it has of its own.
When a person has to step in
Some questions are beyond a branch table: a billing dispute, a lost card, an injury at a gym. Two ways to pass them on:
- Tell the agent in its instructions to take a contact address and promise a call back, then let an Email block send the conversation to your service inbox.
- Or put an AI router before the agent with outputs such as Branch question and Account or complaint. The second goes to Ask a person, which asks your duty manager in Slack or by email and carries on with their answer.
On a chat page the visitor may have left by the time a manager answers, so the first route usually fits better there. On WhatsApp, the answer from Ask a person can go back into the same chat with Reply.
A worked example (made up)
A gym chain in Bangkok has nine branches with different hours on Buddhist holidays, and its front desks spend the mornings answering the same calls. The service lead uploads the branch table with a column for holiday hours, adds the membership terms as a reference file, and builds the flow above with Topics set to branches, opening hours, classes, membership, parking.
Testing in the editor's Chat, she asks for the nearest branch to Siam Square that is open at 6 am on Sunday. The agent compares the two closest, checks their Sunday hours in the table and gives the one that opens at 6 with its phone number. She then asks for a member's home address and gets a polite refusal. Check the work catches one answer that left out a phone number and has it rewritten. She saves eight such exchanges as test cases before switching the chat page on.
Mistakes to avoid
- Keeping hours in the instructions. When a branch changes its hours, update one row, not the prompt.
- Giving the agent no instruction for the unknown. Without one it will try to help, which is how wrong answers start.
- Testing only friendly questions. Try a rude one, an off-topic one and a question in another language.
- Ignoring the Conversations window. Reading a week of real chats shows which questions to add to the table.
The Branch finder and menu chat template is a quick start for a branch-finder, and the AI employees guide covers what else such an employee can take on.
Questions
Can the agent tell a customer which branch is closest?
Yes, when the branches in your table have positions and the agent knows roughly where the customer is. A place name the customer gives is looked up for you.
Will it invent opening hours it does not have?
Its instructions should tell it to answer only from the table, and Check the work can reject any hours or numbers that are not in the data before the answer is shown.
How do I know what customers really ask?
The flow's Conversations window lists the chats, and thumbs up or down under each answer collect feedback you can turn into test cases.
Can it work on WhatsApp as well as the website?
Yes. The same agent can sit behind a WhatsApp Business message event, and Reply answers in the same chat.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates