How a Polish Online Store Reclaimed Its Team's Time with an AI Email Assistant

A Polish online store was handling dozens of repetitive customer enquiries every day — mostly delivery status questions and invoice requests — consuming time that the team could have spent on work that actually required their expertise. We built and deployed a custom, multi-agent AI system integrated directly with Gmail and the BASE e-commerce platform. The system operates on a human-in-the-loop model: it generates ready-to-send draft replies in seconds, and the team member approves or edits the message with a single click — without leaving their inbox. Email handling time dropped by 5×.

The Client and the Challenge

An online retailer shipping orders to customers across Europe, running its storefront on the BASE platform. The customer service team handled post-purchase communication alongside a range of other operational and administrative responsibilities.

The Problem: Repetition That Costs Real Money

Dozens of messages arrived in the company inbox every day. Analysis showed that the vast majority followed a predictable pattern:

“When will my parcel arrive?” — delivery status and estimated arrival questions, each requiring a manual check in the logistics system.
“Could you send me an invoice for my order?” — requiring the team member to locate, download, and attach the document manually.

Handling each of these messages required the same sequence of steps:

Read the email and identify what the customer was asking.
Log into BASE and locate the order.
Check the shipment status in the tracking system.
Write a personalised reply — often in the customer’s language, which differed from the team’s working language.
Download the invoice and attach it to the response.

A single email could take several minutes. In a foreign language: longer. Across dozens of messages per day, this added up to a significant burden that kept the team away from work that genuinely required human judgement.

The status quo wasn’t broken — it was just far too expensive for what it delivered. Every minute spent copying a tracking number was a minute taken from tasks no machine can do.

The Solution

Phase 1 — Discovery: Understand the Process Before You Automate It

Before writing a single line of code, we conducted a detailed interview with the owner and the team. We needed to understand:

How did the workflow actually run? Who handled incoming messages, in what order, and how were reply decisions made?
What commitments had been made to customers? Which delivery timelines were guaranteed? Which exceptions — delays, out-of-stock situations — required escalation to a human?
What data was available in which systems? Integration with BASE (order data, invoices) and the shipment tracking system was a hard requirement for the agent to function.

This phase allowed us to define the boundaries of the system’s autonomy — precisely identifying which messages the agent could handle independently (by generating a draft) and which should reach a team member without any AI-generated suggestion.

Phase 2 — Building the System: Effectiveness Over Complexity

Based on everything gathered in the Discovery phase, we designed and built a custom solution — tailored to this client’s specific systems, workflows, and team habits. Not an off-the-shelf template, but a system built from the ground up.

At its core is an AI agent that understands the context of an incoming message, retrieves the relevant data, and prepares a reply ready to send.

From the business owner’s perspective, the system works like this:

A customer email arrives — a delivery question or an invoice request.
The system checks automatically — shipment status and order details retrieved from BASE.
A draft reply appears in the inbox — in the customer’s language, with the correct information, a tracking link, or an attached invoice.
The team member approves with one click — or makes a quick edit if needed.

What happens under the hood is intentionally invisible — to the team and to the customers.

Workflow

Phase 3 — Deployment: The Invisible Assistant

The core design requirement was that the system should work in the background — the team should not experience it as a new tool requiring training or a change in how they work. Every element of the interface was embedded within Gmail’s native view:

Draft replies appear directly in the inbox — no separate panels, no additional applications.
Gmail labels allow the team to instantly distinguish agent-handled messages from those requiring full human attention.
Zero extra windows — the team member reads the draft, clicks Send, or makes a small edit and sends. Everything happens in one place.

We deliberately chose the human-in-the-loop model over full autonomy. The reasoning was straightforward: in e-commerce customer service, brand reputation is too important to risk sending an incorrect or inappropriate reply without human oversight. The agent is a tool that accelerates work — it does not replace human judgement.

Key Results

5× faster email handling — the agent generates a complete draft in seconds, replacing several minutes of manual work.
Native multilingual support — replies are written in the customer’s language, with no translation time required from the team.
Automatic invoice attachment — documents are retrieved from BASE and attached to replies without any human involvement.
Automatic tracking links — shipment tracking URLs are pulled and inserted into replies by the agent.
A cleaner inbox — Gmail labels let the team instantly filter agent-handled messages from those that need their full attention.
Time returned to the team — customer service staff can now focus on administrative, operational, and complex tasks that require real expertise and experience.

Conclusion

The most common mistake in customer service automation is pushing for full autonomy from day one. Trust in an AI system is built gradually — on both the business owner’s side and the team’s. The human-in-the-loop model we applied gives the agent room to operate while keeping a human in control where the client relationship is at stake.

What also mattered in this project was that we didn’t change the team’s tools. Gmail stayed Gmail. No one had to learn a new application — the agent came to them, not the other way around. That design decision has a direct impact on adoption.

Automating repetitive customer email handling is one of the most measurable and fastest-payback applications of AI in e-commerce. It doesn’t require an IT overhaul — it requires a precise understanding of the process, a well-designed system, and a sensible approach to the boundaries of automation.

If your customer service team is spending a significant part of their day answering the same questions — that time can be reclaimed.

Let’s talk about how to make it work in your business.

📩 Get in touch · 🔗 See other projects