No. For a fashion retailer in Dubai, the AI that now fills its marketplace templates never touched the ERP. It got a plain export, five to ten real products with their finished listings, and an explanation of the team's internal codes. What it really had to learn was the knowledge that lived only in the team's heads.
I have lived and worked in Dubai for four years, and this project came from a retailer here. The products live in an ERP, and the business sells through its own online shop and through the big Gulf marketplaces, each with its own template. Every few weeks new products have to go into all of those templates, and until this project someone on the team filled the columns by hand. Many people assume that automating this means wiring an AI straight into the ERP. In this case the client ruled that out on the audit call, and the project was built without it.
Keeping the ERP closed made the project simpler
The retailer was clear from the start. They would give the AI their data, but they did not want Claude or any other AI tool connected to the ERP. My answer on the call was short: it just makes everything simpler, and it was fine for me.
What the AI works from instead is an export. The team already knew how to produce it as a CSV, so the AI starts from exactly the file a person on the team starts from today. I even raised on the call that we might automate the export itself later, but that was a maybe and not a requirement. Nothing in the business system had to change for the automation to work.
The real work therefore moved to a different question, and it is the one that decides the result: does the AI understand what is inside that file?
The real input is the knowledge in your team's heads
The export was full of abbreviations that only made sense inside the company. A fabric was written as four letters, a product category as two, and the people who fill the templates had learned over the years what each one means. One of them said it plainly on the call: everyone who works with the product understands these short words, because each one has a meaning.
That knowledge is the actual material of the project. I put it to the team like this: "I give it the input. I give it the output. I tell it what it means. And teach it what at the moment only is in your heads."
My first request was one sheet that explains every abbreviation. When the team answered that this list would almost never end, I asked for screenshots of what they already had instead, with no need to write anything down. The AI needs to understand the codes, and a screenshot of the sheet the team already uses carries them just as well.
Five to ten finished products teach more than a written rulebook
The most useful thing a client can hand over is finished work. I asked for five to ten real products, with the columns and values from the export, the completed marketplace and shop templates, and the titles and descriptions that went live. From examples like these I can find the patterns and teach the AI why the team did it that way.
Toward the end of the call the team suggested consolidating and cleaning the files before sending them. I asked them not to. I did not want them to spend a few hours on preparation, and a completely filled sheet in its raw form is exactly where the patterns of their real work show. As I told them, the more data you can give me, the better I can teach it.
Where the examples are not enough, the testing finds it. In a round of testing, if the AI has not understood something, we add a specific rule and tell it: "No, you missed. You made this mistake."
A bad export showed what the skill had learned
The test that convinced me came after the build. I ran the finished skill on a fresh export from the client, and the export was poor: no barcodes and no sizes. The skill noticed. It reported what was missing, pointed out one category that was not in the master list and asked what it should do with it. In the database the result looked really good, and I was genuinely happy with what the skill did. The verdict I waited for after that was the feedback of the person who had filled these templates by hand, because their judgement is the one that counts. I was sure we could make that client happy.
A few weeks later the skill was up and running, and when I looked at the output again the difference was obvious. The product descriptions were all different now, every item with its own text.
What it takes to start is smaller than most people expect
You do not need an ERP integration to automate marketplace listings. You need an export, a handful of finished examples and the codes your team carries in its head, and then test runs until the team is happy with what comes out.
The tool side is modest too. My suggestion to this client was a single Claude account, which costs about 20 dollars a month, set up so that the team can ask for changes to the generated sheet in plain language instead of editing it by hand. For the build itself I aimed for two weeks, and I said why: it is better that it works perfectly than that it ships too quickly, so there would be a lot of testing.
If you sell through the Gulf marketplaces from Dubai, start where this retailer started: collect five to ten products exactly as they went live, together with the raw export they came from.