AI automation for e-commerce companies with margin counted after returns
In an online shop every system tells a different truth: the platform reports sales, the warehouse reports stock, accounting reports finance, and the ad panel reports its own version. We pull them into one place and add an AI agent that answers how much a single product really earns, after returns and after ad spend.
- 14 daysto the first automation running, counted from the end of the workshop
- EUR 3,500net per week of workshop, a published price we do not negotiate
- 3.53 to 4.19how return on ad spend moved in our own shop over six months
Problem
Why is it unclear which product actually earns?
Sales sit in the shop panel, cost of goods in accounting, ad spend in Google and Meta, and returns in a spreadsheet somebody keeps by hand. Each of those places tells the truth and none of them tells the whole of it. Margin per product only exists once somebody lines them up, which means once a month, if at all.
Returns are the most expensive part of that, because they are the worst counted. A product with a return rate above thirty percent can carry a negative margin on positive sales, and in the ad panel it still looks like a bestseller, because the panel counts what was sold, not what stayed sold.
On a marketplace the return mechanics belong to the platform, so the only thing left is work at the source: the size data in the listing and the construction of the garment. To do that you need to know which product comes back on which market, not what the average is.
- Ad budget goes to products that do not earn once returns are counted
- An empty shelf in the size that was selling, because the order was placed from memory
- Customer service answers the same status question for the hundredth time
- Margin per product known after the month closes, too late to act on
Why now
What changed that makes this pay off today
Shop platforms, warehouse systems and ad panels now expose interfaces you can read data through without exporting to a spreadsheet. Three years ago wiring that together was a quarter-long project; today it is counted in working days. The automation layer is documented in public, n8n and Make both publish what their connectors do, so nobody has to guess.
The second change matters more. A rule will move a number from one system to another, but it will not answer the question why margin in Germany dropped last week. A language model reads sales, returns, cost of goods and ad spend at once and answers in a sentence instead of another export.
The third one is market. At current customer acquisition costs, three points of difference in the return rate stopped being a statistic and became the line between a campaign that pays and one that burns. That is exactly the level of detail a spreadsheet will not hold.
What changes
What do you see after the first month?
Margin after returns, not before
Sales, cost of goods, ad spend and returns lined up per product and per market. You see what earns, rather than what looks good in a panel.
A daily figure instead of one a month, after the books close
Returns counted per product
Which style, size and market comes back most often, and for what reason. That is a list of things to fix in the listing, not an average for a report.
Split by product and market, not one number for the whole shop
Advertising cut from data
Products that do not earn after returns drop out of campaigns before they finish the budget. A rule on data makes that call, not a hunch.
302 products cut from campaigns in our own shop over six months
Stock-outs visible before the stock-out
A forecast built on rotation and delivery lead times shows what runs out in two weeks, instead of what ran out yesterday.
Purchasing based on consumption, not on memory
Order status without asking support
Questions about an order, a shipment or a return have an answer in the system, so support handles exceptions instead of repeats.
16 people in our own shop ask the agent every day, the warehouse included
Proof
Do you use this yourselves?
Yes, and it is the only proof we can show without asking a client for permission to publish their figures. Two AI agents running in production inside our founders' own companies, both launched in March 2026 and built by the same method described above. Below is six months of measurement, together with what these numbers do NOT prove.
Built for ourselves. These are not client deployments but implementations inside the founders' own companies. We say so plainly, because proof you cannot verify is not proof.
JustBottheir agentJustIdea Agency, a performance marketing agency of around 30 people
- 23
- people from the company work with the agent
- 38
- systems wired in permanently
- 112
- processes run without a human
- 475
- thousand lines of the agent’s code
An agent wired into Google Ads, Meta, the CRM, the task system, the SEO tool, Analytics, Search Console and the data warehouse. First commit 1 March 2026, today 23 people use it.
8 of 38 systems it works in
Google Ads
Meta Ads- HubSpot
- ActiveCollab
- Ahrefs
GA4
Search Console
BigQuery
- Before
- In the twelve months before launch the agency ran at a margin of minus 4.42 percent. Across a twenty-month window there was not a single run of three profitable months.
- After
- May, June and July 2026 are three profitable months in a row, the first such run in the whole window. Revenue barely moved, so the change sits on the cost of delivery, not on sales.
Margin up 5.55 percentage points year on year, with revenue up 1.5 percent
SurfBottheir agentSurf Inc., a premium clothing brand with an online shop and two physical stores
- 16
- people from the company work with the agent
- 21
- systems wired in permanently
- 89
- processes run without a human
- 323
- thousand lines of the agent’s code
An agent wired into the order system, the ERP, the shop, the data warehouse and the ad platforms. Launched in March 2026, fully running by May. Sixteen people use it, warehouse and store staff included.
8 of 21 systems it works in
PrestaShop- BaseLinker
- Comarch Optima
- BigQuery
- Google Ads
- Meta Ads
- GA4
Zalando
- Before
- From February to August 2025 every unit of ad spend came back as 3.53 in online revenue, and advertising absorbed 28.3 percent of that revenue.
- After
- In the same months of 2026 the ratio is 4.19 and advertising is down to 23.9 percent of revenue. The ad budget is 6.9 percent lower with sales 10.3 percent higher.
Return on ad spend from 3.53 to 4.19, 18.5 percent higher, on a budget 6.9 percent lower
2026, with the agent 2025, before the agent revenue per unit of spend
Return on every unit of ad spend at the clothing brand, February to August, 2026 against 2025. The 2026 curve starts LOWER than the year before, breaks away in March and holds the lead for the rest of the season. The peak is 4.91 in May. We point out the weaker start because it shows we did not pick a convenient baseline.
Measurement: Online revenue from the order system against spend across every ad platform, the same months in both years. Cost data in the warehouse starts in February 2025, so we cannot rule out that the ratio was already rising on its own.
Operations run by the agency agent in successive months of 2026, from the first commit in March. From 83 operations in March to almost 63 thousand in August, with users growing from 4 to 23. This curve is here because a deployment nobody uses looks identical in a report to a successful one.
Measurement: Agent call log in the data warehouse, March to August 2026, counted at the level of a single tool invocation.
Breakdown of 29,854 operations run without a human over thirty days, by kind of task. Most of it goes to tracking team availability and monitoring client sites: things that have to run without a break and that nobody wants to do by hand.
Measurement: Agent audit log, the 30 days before 2 September 2026, counted by the kind of tool invoked.
Of 64,903 operations run in the last thirty days, 29,854 went through without a human: client site monitoring, warehouse loading, ad account reviews, catching form submissions. The rest is work somebody requested and reviewed. We give both numbers, because autonomy without oversight is not a feature but a risk.
Measurement: Agent audit log, the 30 days before 2 September 2026, split by whether a call had an associated human session.
What about you
What if our shop works differently?
Four sentences come up in shops more often than any others, and all four are fair.
We run PrestaShop, not Shopify, so this probably will not work.
It will, and this is the case we know best: our own shop runs on PrestaShop wired into an order system and an ERP. We read the data through the interface and we leave the platform where it is.
We sell on several marketplaces and each one holds its own data.
That is usually a reason to start, not to postpone. The gap between channels costs most exactly when you need to compare margin after returns, and every channel counts a return differently. Bringing that into one table is the first thing a rollout delivers.
We have no IT department, nobody here can look after it.
The technical part is ours. After the rollout it is run by the person who knows the shop, not by a developer. In our own shop the warehouse and the physical store staff use the agent.
Our margin data is sensitive.
We work on your infrastructure or on a cloud you name, and we agree the scope of access during the workshop and write it down. It is the same arrangement we keep our own sales data under.
- Shop platforms
- PrestaShop, Shopify, WooCommerce, Shopware, Magento
- Orders and marketplaces
- BaseLinker, Allegro, Amazon, Zalando, eBay
- Warehouse and logistics
- WMS, couriers, pick and pack systems, delivery booking
- Accounting and ERP
- Comarch Optima, enova365, Subiekt, Fakturownia, Xentral
- Advertising and data
- Google Ads, Meta Ads, GA4, BigQuery, Google Sheets
- Automation layer
- Make, n8n, custom integrations over API
What exactly we automate in a shop
A typical split of the work in an online shop before a rollout and after it.
| Area | Today | After the rollout |
|---|---|---|
| Margin per product | assembled by hand after the month closes | counted daily, after returns and after ad spend |
| Returns | one average for the whole shop | split by product, size and market, with a reason |
| Ad budget | products cut on a hunch | a rule cuts what does not earn after returns |
| Stock and purchasing | run from memory and a spreadsheet | stock-out forecast from rotation and lead times |
| Order status | a question to support | an answer from the system, support takes the exceptions |
| Marketplace listing | filled in when somebody spots a gap | a nightly list of gaps that block a listing |
We agree the scope of the first rollout during the workshop. We take the three items with the highest return, not all of them at once.
How we work
Four steps from spreadsheet to agent
Each step ends with something you can judge before you decide on the next one.
Workshop
One week. We walk an order from the click to the return and the settlement, and check where the data drifts apart. We leave with a map and a list of places where the same number is produced twice.
Picking three automations
From the map we pick the three with the highest return and measure them before we start, so that after thirty days there is a way to tell whether they worked.
First rollout
Fourteen days to something running. In a shop we usually start with margin after returns, because that is what changes budget decisions in the first week.
Measurement, then the next one
After thirty and ninety days we check the indicators agreed during the workshop. Only then do we decide on the next area.
Before and after
How the week of the person running the shop changes
Before
- Monday goes on gluing data together from four panels
- The budget call is made on sales, because returns land later
- A missing size shows up when a customer cannot buy it
- Margin per product is a project, not a number you check
After
- Monday starts with the figures already assembled, ready to comment on
- Budget follows margin after returns, not sales on their own
- The forecast shows the gap before the shelf empties
- Margin per product is a question asked in a plain sentence
The amount of data does not change. What changes is whether you can make a decision from it the same day.
For the owner and the head of e-commerce
What the person ordering this gets out of it
The first thing shows up straight away: the conversation about results stops opening with where this number came from and starts with what to do about it.
The second is less obvious and matters more in front of an investor or a bank. Profitability of a single product after returns stops being one person's knowledge and becomes a figure in a system you can show and defend.
What it costs
The workshop price is published, the rollout is quoted after it
The workshop costs EUR 3,500 net per week and runs from 1 to 7 weeks. We give that price before the first call, because a shop has a right to know what it is agreeing to. We quote the rollout after the workshop, once we know how many systems we are connecting and what state they are in. Tool and model costs you pay straight to the vendor, at EUR 0.005 to 0.07 per query at provider list prices, with no margin from us.
Process workshop
EUR 3,500 net
per week, usually one to three
A closed piece of work. It ends with a process map and a 30 day plan you can run without us.
- A week inside the shop and inside the systems
- A map of the order road from click to return
- Three automations with the highest return, measured
- Indicators to check after 30 and 90 days
- Most popular
Workshop and first rollout
quoted after the workshop
For a shop that wants margin after returns counted within two weeks of the workshop ending.
- Everything from the workshop
- First automation running in 14 days
- Connection to the platform, the warehouse and the ad panels
- Training for the person who runs it on your side
- Measurement after 30 and 90 days
AI agent for the shop
quoted after the workshop
For a shop where questions about sales, stock and returns come up a dozen times a day.
- Everything from the rollout
- An agent answering questions on sales, margin, stock and returns
- Alerts when margin or the return rate drifts
- Access for the team, the warehouse and customer service
If the workshop shows there is nothing worth automating, we say so and we do not propose a rollout. It has happened and it will happen again.
Frequently asked questions
What is e-commerce automation?
It is a system taking over the work somebody does by hand today: lining sales up against cost of goods and ad spend, counting returns per product, watching stock levels and answering questions about order status. The data for that is usually already in the shop, it just sits in four systems at once. Automation means bringing it into one place, not buying another platform.
Do we have to change our shop platform?
No. We read the data through the interface and leave the platform where it is. Our own shop, the one we tested this on, runs on PrestaShop. A platform migration is a quarter-long project, while the first automation is supposed to be running in fourteen days, so the two do not belong in the same plan.
How do you count margin after returns?
We line net sales after returns up against cost of goods from accounting and ad spend from the panels, per product and per market. The last part is the one that matters: a return is charged to the product and the market it came from, not to a shop average. Without that, a product with a return rate above thirty percent looks profitable, because its sales figure is positive.
Will this work when we sell on Amazon and Allegro?
Yes, and it is usually worth more there than in your own shop, because the return mechanics belong to the platform and cannot be changed. What is left is work at the source: size data in the listing and the construction of the product. To do that you need to know which product comes back on which market, and that is a number the marketplace report will not give you.
How long from signing to the first result?
One week of workshop and fourteen days for the first rollout, counted from the end of it. We measure the effect after thirty and ninety days, on indicators agreed before the start, so that nobody judges it on impressions. If the numbers do not move, that is visible on the same indicators, which is the point of setting them first.
What does the rollout cost after the workshop?
It depends on how many systems we connect and whether the data inside them is in order. We quote it after the workshop, because that is the first moment we know the answer. The workshop itself costs EUR 3,500 net per week and you know that price before the first call. Model costs run at EUR 0.005 to 0.07 per query and go straight to the provider.
Let us start with one question
In the free 30 minute consultation we walk one of your processes, usually the road an order takes from the click to the return, and tell you whether we see anything there worth automating. If we do not see it, we say so on the call.
We run two workshops a month, because a week inside a company is a week of one person's work on our side.
Book a free consultationPS. If you want to see the numbers from our own shop first, they are further up this page, together with what they do not prove.