Marcin Dobrzyński, home page

Case studies: e-commerce and AI in practice

Seven stories from my own companies, each told the same way: the situation, what I did, the result. At the end: all figures with definitions and the method.

01 · Legal form

From a Polish company through an Einzelunternehmen to a GmbH in Germany

2 formsEinzelunternehmen and GmbH in Germany

Situation

Since 2011 I sold to Germany and Austria as a Polish company. As sales grew, a presence in Germany became necessary: first a sole proprietorship (Einzelunternehmen), later a GmbH.

What I did

I registered an Einzelunternehmen and sold from within Germany as well, then carried out the operational side for the GmbH: Amazon and eBay accounts, separate Billbee accounts, VerpackG and WEEE registrations, and invoicing and document flow for VAT, intra-EU supplies and OSS, implemented as agreed with the accountant and Steuerberater. I am the GmbH’s managing director (Geschäftsführer).

Result

Today three companies sell from two warehouses, in Poland and in Germany, each with its own accounting and accounts, on a shared stock synchronisation. I know both forms in Germany from registration to day-to-day operations.

This is a practitioner’s account, not legal or tax advice. Make decisions about legal form and taxes with a tax adviser (in Germany: Steuerberater) and a lawyer.

Tools: Amazon Seller Central, eBay, Billbee (multiple accounts), iFirma, LUCID, in-house stock synchronisation.

02 · Amazon.de

Amazon.de: buyable listings and advertising kept in line with margin

9 in 10listings published by the system are buyable

Situation

A large Amazon DE account with an extensive catalogue. Some listings could not be bought, and it was hard to tell whether the advertising was making money or just generating turnover.

What I did

My system publishes the listings: with complete attributes, GPSR data and listing analytics on historical data (including Keepa). I judge advertising by margin and set bids and budgets from the break-even point.

Result

More than 9 in 10 listings published by the system are buyable on Amazon.de (91.6%, as of 25.09.2026). Starting point: old imported listings outside the system were often not buyable because of GPSR blocks and SKU aliases without stock sync; I clean those up separately.

Tools: Amazon SP-API, Amazon Ads, Keepa, net margin analysis, Claude Code.

03 · Marketplaces in Germany

A second marketplace account in Germany: same system, new pace

3×more orders year on year, Q3 2025 → Q3 2026 (to 25 September)

Situation

The second seller account in Germany ran alongside the main one, but by hand: listings created one by one, stock tracked separately, advertising without shared logic.

What I did

I connected the account to my system: shared stock across both warehouses, automatic listing on eBay and Amazon, margin-based repricing and advertising.

Result

In Q3 2026 (to 25 September) the account has three times as many orders as a year earlier: eBay ×4.5, Amazon ×2.5. One seller account belongs to one company; I count orders from all channels of that account. Part of the growth comes from moving listings over from the main account.

Tools: eBay API, Amazon SP-API, Billbee, in-house stock synchronisation, Claude Code.

04 · Allegro

Allegro: a new channel with automatic listing

4×from 22 to 92 orders, Q2 → Q3 2026 (to 25 September)

Situation

There was stock in the warehouse worth listing. Preparing each listing by hand (title, description, attributes, photos) kept losing out to more urgent work.

What I did

An automated process prepares the listings and emails them for approval. One click approves a listing, and the link is cryptographically signed, so it cannot be forged. A separate flow covers small items, with the mark-up calculated automatically.

Result

The channel launched in 2026 had 22 orders in Q2 and 92 in Q3 (to 25 September).

Tools: Allegro via Apilo, HMAC-signed approval links, content generation with Claude.

05 · Advertising

Advertising judged by margin, not revenue

−53%eBay advertising cost for the same sales attributed to ads

Situation

The ad budget was spread across several channels, and each one reported success in its own metrics.

What I did

I brought advertising results into one view, measured by net profit rather than ad revenue. On top of that an automated ad adviser proposes bid changes every day for approval, and I set eBay campaigns by margin.

Result

On eBay the same sales attributed to ads cost 53% less (June → August 2026). On Amazon

Tools: eBay Marketing API, Amazon Ads API, Google Ads API, Claude Code.

06 · Operations

Three companies, two warehouses and one system

3 · 2 · 1companies, warehouses, system

Situation

Three companies from Poland and Germany sell from two warehouses, in Poland and in Germany, on shared channels. Every company can buy and ship from both, and a transport runs between the warehouses every day. Three sets of books, three sets of accounts. Without a shared source of truth, the same item can be “in stock” in three places at once.

What I did

One source of truth for stock in both warehouses and a layer that knows which company a listing belongs to and how to settle a sale. Further tools sit on top of it: my own auto-listers running alongside Magnalister, a print queue for invoices and labels and, since August 2026, my own AI helpdesk instead of LiveAgent, including a voice bot on the phone line (sipgate) that opens tickets and recognises customers by number.

Result

A sale on any channel deducts stock for every company and in both warehouses. Each company has its own accounts, permissions and documents; the system shares only warehouse stock and logic. The tools know the business context, so they do more than off-the-shelf subscriptions; where a ready-made tool works well, it stays and is connected better.

Tools: Claude Code, Python and FastAPI, PostgreSQL, an event queue, marketplace integrations via official APIs.

07 · Gambio websites

Gambio websites: modules that did not exist, written with AI

Gambiocustom modules and listing templates

Situation

My Gambio websites had standard product pages and a standard basket. What raises order value was missing: templates matched to the product, sensible cross-selling and a clear indication of how much is left to free shipping.

What I did

I wrote my own Gambio modules myself, with Claude, and deployed them on my websites, including an interactive basket bar that shows live how much is left to free shipping. My system has access to these channels, so it generates listing templates per product and cross- and upselling automatically from catalogue data.

Result

Listings are more attractive without manual work on every product page, and customers see in the basket what to add to get free shipping.

Tools: Gambio (custom modules), Claude Code, catalogue data from my system.

All figures with definitions

As of 25.09.2026. Q3 2026 is counted to 25.09, so it is incomplete. Periods and sources are described below in the method.

  1. 1.2 h

    Median time to the first human reply in the helpdesk, June → September 2026 (to 25 September): from 9.7 to 1.2 h with five times more conversations. Clock hours, excluding automatic acknowledgements. Every second reply starts from an AI draft.

  2. −53%

    eBay advertising cost for the same sales attributed to ads, June → August 2026, after rebuilding the campaigns around margin.

  3. 9,000+

    New listings published automatically since February 2026 (to 25 September): Amazon c. 3,400, eBay c. 4,600 auctions, Allegro c. 1,100, Kaufland c. 270. Distinct new listings, excluding imports and updates of existing ones.

  4. c. 230

    Supplier orders prepared by the system and sent via API in 4 weeks (31 August to 25 September 2026). A person approves the quantities.

  5. 11×

    Orders on domabau.pl: from 12 to 132 per quarter, Q1 → Q3 2026 (to 25 September). The website started from zero in January 2026.

  6. 4×

    Orders on Allegro: from 22 to 92, Q2 → Q3 2026 (to 25 September), with listings published automatically.

  7. 3×

    Orders on the second seller account in Germany, Q3 2025 → Q3 2026 (to 25 September): eBay ×4.5, Amazon ×2.5. One seller account belongs to one company; part of the growth comes from moving listings over from the main account.

  8. 5% → 19%

    Share of orders from channels developed since 2025 outside the main account (second account in Germany, domabau.pl, Allegro), Q3 2025 → Q3 2026 (to 25 September). Today one order in five.

  9. 91.6%

    Buyable listings published by my system on Amazon.de, as of 25.09.2026. Starting point: old imported listings outside the system were often not buyable because of GPSR blocks and SKU aliases.

  10. +83%

    Amazon advertising budget, scaled between Q1 and Q2 2026 while keeping ROAS above break-even in every month of 2026 (ad sales ×2). This is scaling, not higher efficiency.

  11. 95%

    Requests and notifications closed without a person, out of more than 41,000 since August 2026 (to 25 September), since my own helpdesk replaced LiveAgent. Mostly automatic messages from marketplaces and systems; customer matters go to people.

  12. 2,600

    Invoices a month print themselves (August 2026). The manual printing and matching of documents to parcels is gone.

  13. 1,600–2,100 h

    Estimate of manual work taken off the team since January 2026. Assumptions: 8–10 min per listing, 3 min per reply, a few seconds per request closed automatically. This converts workload into hours; it is not a recovered job.

Method

Facing a similar situation?

Tell me what it is about. I will reply whether I have experience with it.

Get in touch →Tool catalogue