Case study
dezShop: an AI agent that takes customer requests on WhatsApp
An automatic message after a missed call, a model that decides who replies, and an agent that asks for missing details and opens the request in the CRM.
dezShop sells used car parts and is part of our own operations. The team receives more than 1,000 WhatsApp messages a day, and many calls go unanswered at busy times. azai built the flows that take over part of these requests automatically, on the same number the operators use.
The problem
A parts request needs the car, the engine and the exact part. Customers usually write only part of it, and the operator exchanges several messages to find out the rest. Meanwhile, missed calls go unanswered and the customer looks elsewhere.
How the data moves
- 01
Missed call
The phone system sends the missed call to the CRM. If nobody has picked it up within 8 minutes, the customer receives a WhatsApp message from a template approved by Meta.
- 02
Choosing who replies
When the customer writes, a small model reads the conversation and picks one of three routes: the agent takes it, it is only a confirmation, or a person is needed.
- 03
The agent
The agent asks for what is missing, reads the photos sent and opens the request with the parts and the car. When the request is complete, it moves it to the colleague who sets the price.
- 04
The team
A person gives the price and the quote. Conversations the agent does not take stay unread in the inbox, as before.
What the system checks
People come first
If a colleague has written in the conversation in the last 15 minutes, the agent stays out. If a colleague replies while the agent is working, the agent's message is rejected.
Limits and hours
Each flow has a daily limit and a time window. The missed-call message goes to the same number at most once a day.
A log for every decision
Every routing decision and every agent message stays in the log, with the reason. From the first day's log we corrected the rules for cases wrongly sent to a person.
How one request moved through the CRM
The steps of a real request from 10 October 2026. We show what changed in the CRM, not the customer's messages.
| Time | What happened | What changed in the CRM |
|---|---|---|
| 10:04 | Missed call | New request under “Missed calls” |
| 10:12 | Automatic WhatsApp message | The request waits for the customer's reply |
| 10:20–10:24 | The customer lists the parts; the agent asks which radiators | Parts added, status “Waiting for customer details” |
| 10:27–10:31 | The customer confirms; the agent asks for a photo of the registration document | The generic radiator replaced by three parts |
| 10:33–10:37 | The customer sends the document; the agent confirms | Car details filled in, request moves to pricing |
The result, and applying it elsewhere
On the agent's first day, 10 October 2026, the routing model gave the agent 41 conversations, left 29 to the team and marked 12 as simple confirmations. The agent sent 70 messages, with a median reply time of 4 minutes after the customer's message. The missed-call message went out 2,889 times in the first 30 days, and 58% of customers replied on WhatsApp within 48 hours.
The example comes from car parts, but the flow fits any business that receives requests on WhatsApp: repair shops, distributors, rentals, installers. For another company we write different rules and connect the agent to the existing CRM or spreadsheet. A voice agent that answers the phone is built on request.