Catalog agents: merchant onboarding in minutes, not days
The problem
A new merchant arrives with a menu or product catalog exported from the system they're leaving: Toast, Square, Clover. The export never fits ours. Item names, modifiers, categories and prices are structured differently, and someone on the onboarding team reshapes it by hand. By my count that took three to ten business days per merchant, and it sits at the moment a new customer is most likely to walk away. When I started looking, the complaints were specific: the error messages were tedious, and the upload process was clunky.
What I saw
I recorded the meeting where the team described their problems, and I went back to the recording again and again until I understood what they were going through. That's what got me past the error messages. Once I felt the pain, the question changed from how to fix the errors to how to make onboarding easy in the first place.
A restaurant menu, a service price list and a retail catalog break in different ways. Restaurants have modifiers and combos. Service businesses have durations and staff-specific pricing. Retail has variants and SKUs. One general-purpose converter would be mediocre at all three, so I treated each vertical as its own problem.
The target is strict, too. talech takes a catalog as an Excel template with separate tabs for products, categories, tags, suppliers, modifiers and modifier pricing. Names are case sensitive, so "hat" and "Hat" become two categories. A category name is capped at 40 characters and a product name at 100. A product can sit in only one category. Variations and modifiers have to be built on their own tabs, and the upload fails on any bad cell. That's all in talech's own bulk upload guide. It's exactly the kind of rule-heavy translation where a person loses hours and a model, with a reviewer, does well.
What I built
Three Microsoft Copilot Studio agents, one each for restaurant, services and retail, each tuned to the anomalies of its vertical. They take the messy export and produce a clean catalog in our upload format, and an onboarding analyst reviews it before it goes in. I thought about those complaints over a weekend, built the first version on a Saturday, and demonstrated it to my colleagues that Monday. All of it was done in Microsoft Copilot Studio, on my own, with no engineering team. They run inside the Bank's existing tenant on existing licensing: no new vendor and no new build budget. The data is catalog and menu data only, with no personal or sensitive customer data.
The agents are in production, and the onboarding team uses them on live merchant data. Ownership now sits with that team, and they tune the agents constantly without me. As the toolset matures, other groups, including customers and prospects, can use it too. The next step on the roadmap was the same treatment for customer data, which customer service was still migrating row by row.
The result
Onboarding took three to ten business days per merchant. It now takes the team three to ten minutes. That's my own measurement from testing, not an audited figure, and I don't have a before and after on merchant churn.
What I learned
Prove it on real data before asking anyone for budget. Three narrow agents beat one clever one, and a human reviewer keeps the speed from costing trust. Handing it to the team that lives with the work mattered as much as building it. And lean in when people bring complaints. Here the complaints were the brief. Record the meeting where people tell you about them, and replay it.