Add an AI shop assistant and buyers who answer questions

Put the agents in the chamber

Start with a shop such as the one in Build a shopping study with a budget and reviews.

First, add the agents to the chamber of the shop, as for all chamber agents. Here, Nora is the assistant and Marta answers as a buyer.

• Each agent needs a trigger that is active in the shop segment. This is usually an LLM trigger.
• The experiment needs a key to call the model: its own API key, or platform credit.

The Store Builder shows the agents of the chamber by name. An agent without a role in the shop does not speak in the shop.

A buyer who answers questions

On the Reviews tab:

• Add a reviewer.
• Give it the name that participants see.
• Change it from Scripted to Agent.
• In Written by, select a chamber agent by name. You can use each agent one time in each shop.
• Select the products that it answers questions about. If you select no products, it answers about all products.
• Set a reply delay in seconds.

Carrier holds the answer until the delay ends. Thus the answer does not come faster than a person can type.

An agent reviewer only answers a question or a reply to a review. It never writes first.

When a reviewer is an agent, the builder shows a warning. Participants see a name and "Verified purchase". They never see that an agent wrote the text. Make sure that your ethics approval includes this, and plan a debrief.

A shop assistant

On the Assistant tab:

• Turn on the assistant.
• Select the agent that answers.
• Give the assistant the name that shows on its button and drawer.
• Add suggested first messages. They show as buttons until the participant writes.

Before each reply, Carrier adds the assistant's information to its prompt. This is the catalogue, the product in view, the cart and, if you select it, the reviews.

The participant's page sends only product ids. All names and prices come from your store. Thus text that a participant types or changes in the page cannot change them.

The assistant's greeting and timed messages show in its drawer. To send a greeting, give Nora a salute trigger or a time trigger.

Save the store and deploy.

What the participant sees

The participant adds the sport earbuds to the cart, opens Nora and taps the suggestion. Nora answers with the prices from the store.

On the product page, the participant asks if the earbuds stay in during running. The question shows "No answers yet" for 3 seconds. Then Marta's answer shows under Marta K., with Verified purchase.

Each agent sees only its own conversation. Nora never sees product questions. Marta never sees the assistant chat, or questions about products that she does not answer.

What is recorded

The two conversations are in the action log of the shop. The export shows it in Shop Actions, with one row for each action:

• question_sent and reply_sent: what the participant asked.
• answer_shown: when an answer showed (after its delay), with the name of the reviewer.
• assistant_open, assistant_close, assistant_message and assistant_reply_shown: the assistant actions.

The messages are also in Chat History.

The shop variable _assistantMessages counts the participant's messages to the assistant. You can use it in later segments, conditions and prompts.