Example build, fictional company · Esimerkkikauppa
Web-shop customer service agent: answers only from the shop's own data
The customer service agent gets a ready draft with its sources and sees at once which messages need a person.
A fictional web shop's customer service gets questions whose answers already sit in the order data or the delivery terms. In the example the build was given the shop's terms, 12 orders and the customers' messages.
Customer messages
11
Answered from sources
6
Handed to a person
5
Pick a message or a filter.
Customer service agent for email · Esimerkkikauppa
Incoming message
Laura Kuusi · Where is my order?
Hi, last Friday I ordered mugs and a cloth (order 4821), and the parcel has not turned up yet. Has it been sent? Laura
Sources the agent used
- Order 4821
- Terms: Delivery
What the agent found
- Order
- 4821 (dispatched)
- Summary
- The customer asks whether order 4821 (mugs and a cloth) has been sent. The order was dispatched on 29 September 2026 to a Posti pickup point, and the tracking code is known.
- Notes
- The order is under the customer's own email address.
- The last sentence, about a separate email with the tracking code, is based on the terms, not on a confirmed sending. Remove it if needed.
Task
No task.
Reply draft for the customer
Hi Laura, thank you for your message. Your order 4821 left the warehouse on 29 September 2026, and it will be delivered to a Posti pickup point. Your tracking code is JJFI64000012345600021. The delivery time from dispatch is 2 to 4 business days, so the parcel should reach the pickup point within that time. We also sent the tracking code in a separate email when the order was dispatched. Kind regards Esimerkkikauppa customer service asiakaspalvelu@esimerkkikauppa.example This message was drafted with the help of an AI assistant and reviewed by a member of our staff before sending.
A recorded run from 30 September 2026, made with the real language model on fictional data. The drafts are shown as they are left for a person to review. Company names, addresses and phone numbers were changed after the run so that they do not resemble real ones.
What was done
- The program looked up the customer's own orders before the language model saw the message.
- The language model wrote the reply from the shop's terms and order data and named its sources.
- Messages the sources could not answer, or where the decision belongs to a person, went to a customer service agent.
- The query about another person's order was recognised, and the order's details were not disclosed.
What the build did not do
- Sent no replies. The example runs in draft mode.
- Granted no refunds and changed no orders.
- Did not answer the stock question, because stock levels were not in the sources.
- Was not connected to a web-shop platform: the orders were read from a CSV file.
This example is built on
Customer service agent for email