All cases

Case 01 — 2026

Support Assistant

AI agent for e-commerce

A support team of 11 was drowning in “where is my order” and “how do I return this”. We built an agent that answers on its own, pulls real data from the CRM and hands a person everything outside its scope.

Client
Online fashion retailer, 28,000 orders a month
Industry
E-commerce
Timeline
6 weeks to production
Team
2 engineers, a product owner on the client side

Outcome

−47%
of tickets reach an operator
1.8 s
average reply instead of 40 minutes
24/7
chat answers nights and weekends

The task

At peak season the chat queue reached 40 minutes. Seven out of ten tickets were routine: delivery status, return windows, sizing. Operators answered by hand, assembling data from three systems.

  • 40 minutes of waiting at peak season
  • 68% of tickets were the same four scenarios
  • Order data lives in the CRM, the warehouse and the carrier
  • Nights and weekends had no coverage at all

What we built

We started by reading 4,000 conversations and isolating the scenarios where the answer is fully determined by data. Those went to the agent. Everything else deliberately stayed with people.

  • One agent behind both the website chat and Telegram, with shared history
  • Direct calls into the CRM, the warehouse system and the carrier API
  • Returns: the agent files the request and sends the shipping label
  • Hand-off to an operator with a summary and the data already fetched
  • A panel where support reviews the agent's answers and edits a rule

How it works

  1. 01

    The customer writes

    Messages from the website chat and from Telegram reach the same handler.

  2. 02

    The agent picks a scenario

    It recognises the intent and checks whether it has enough data to answer precisely.

  3. 03

    Systems are queried

    Order, delivery status and return terms come from the CRM and the carrier API, not from the model's memory.

  4. 04

    Answer or hand-off

    When confidence is low or the customer is upset, the chat goes to an operator with full context.

Stack

  • OpenAI
  • Anthropic
  • FastAPI
  • Postgres
  • Redis
  • CRM API

“The agent closed half of our routine tickets in the first month, and the team finally works on the hard cases.”

Head of Support

What's next

We are now adding proactive notices: the agent messages the customer about a delayed delivery before they ask.

A similar task?

We'll build a prototype for your process in one day. Free, with no strings attached.

Next caseLegal Copilot