Most companies in North Macedonia use AI one way: a chat window. You paste something in and copy the answer out. The AI never touches your systems, so it only knows what you remembered to give it.
An agent is different. It gets its own access, with rules, and does the digging itself.
What the agent found
A tools retailer runs on Collab, the ERP from Accent Computers, and I'm building their new web shop on top of it. Early in the build I connected an agent to the Collab API to read, and change nothing, and asked it what was actually in there.
Three answers stood out:
- Collab holds 8,106 articles. The old web shop showed 232 products, about 2.9% of that.
- Not one of the old shop's 231 product codes exists in Collab. The old site used manufacturer part numbers, seven to fourteen Latin characters. Collab uses short internal codes, one to five characters, almost half of them with Cyrillic letters. With no key to join the two, the old catalogue could not be carried over, so the new one comes straight from Collab.
- Of 8,128 rows in the barcode table, 49 carry a barcode.
Each of those is one question to the data. By hand, each means exports and spreadsheets, which is why questions like these rarely get asked. Twelve days later the agent ran the same survey again and reported only what had moved.
From my keyboard to the owner's
All of that happened while I was building. The agent worked for me.
MCP, the Model Context Protocol, is how that access gets handed to the people who run the shop. An MCP server sits between Collab and an assistant such as Claude or ChatGPT and offers it a fixed set of tools: look up an article, read the stock for a warehouse, list the promotions. The owner types a question, and the assistant picks the tools it needs to answer it.
I'm building that server for Collab now. Collab's API answers in 120 to 140 milliseconds, so the ERP is never the slow part.
Questions the owner will be able to ask
Every one of these uses data the shop already holds or already reads from Collab:
- How much of the Collab catalogue is online today, and which articles are still missing a photo or a category?
- Which products are down to their last units?
- Which prices in the shop are behind Collab right now?
- Which promotions end this week?
- Which web orders went through without a confirmed stock check?
- What sold online this week, by brand?
- Did the stock sync run every five minutes today, or did it go quiet?
None of them needs a new report or an export. And the next question can build on the last answer: "which of those have stock right now?" A fixed report can't do that.
There are limits. This integration reads articles, stock, prices, promotions and customers. Sales over the counter in the physical shops are not part of it yet, so those questions come later.
Rules before answers
The server is also where the permissions live. It starts read-only. Customer records are a separate switch, off until the owner decides who may see them. Every question and every lookup is logged, so there is a record of what the assistant looked at.
Writes come one at a time, the way the shop itself was built: the only write in its design is a web order, in a document type the client chose.
If you run Collab too
Accent counts over 400 Collab installations. If yours is one of them, the same questions are sitting in your data. This is the work I do under AI agents and MCP integrations.