Projects · E-commerce education · 1M+ community · Microsoft stack
The reporting brain: from ten VAs to one proactive system
An e-commerce education platform with a community of more than a million people replaced hand-built reporting with one enterprise-grade brain on Microsoft Fabric.
The situation
The company runs e-commerce brands and teaches a community of over a million people. Every brand produced numbers in different places: Sellerboard for profitability, the Amazon Selling Partner API for orders and inventory, ad platforms, spreadsheets.
A small army of virtual assistants pulled those numbers together. Five to ten people, every week, exporting, cleaning, merging, and pasting into reports.
The reports told you what already happened. Nobody had time to say what should happen next. And the whole company already lived in the Microsoft ecosystem, so the answer had to live there too.
The week, before
- VAs export data from Sellerboard, brand by brand.
- More exports from the Amazon SP-API and ad consoles.
- Numbers get cleaned and merged in spreadsheets by hand.
- Reports get formatted and posted for the brand teams.
- Managers read yesterday’s news and decide what to check next.
What we deployed
We built one data platform on Microsoft Fabric. Pipelines pull from Sellerboard, the SP-API, and the ad sources on a schedule, clean everything against agreed rules, and land it in one governed model.
On top of the model sits an intelligence layer built with Azure AI Foundry. It does not just chart the numbers. It reads them.
Every morning the system flags what actually needs a human: a product about to run out of stock, a margin sliding under target, ad spend drifting away from plan. Brand directors open a short brief, not a hundred rows.
Sellerboard + SP-API → Fabric pipelines → One governed model → Proactive briefs
How it runs
A run starts on schedule, not when a VA remembers. The pipelines land fresh data from every source into the model, and every number keeps its lineage, so anyone can trace a figure back to where it came from.
Then the agent layer goes to work. It compares every brand’s numbers against the rules the operators set: stock cover targets, margin floors, spend ceilings, velocity changes worth knowing about. Most numbers are fine, and it says nothing about them. That silence is the feature.
What survives the rules becomes the morning brief: a short, ranked list per brand of what deserves attention today, each item with the number behind it and a suggested next step. The brand director reads it in two minutes.
When something is ambiguous, the system does not guess. It holds the item with a note and waits for a person. The team changes the rules whenever they want, in plain language, and every change is visible.
The result is a different shape of week. Nobody assembles anything. The conversation moved from “what happened?” to “here is what we are doing about it.”
What changed
- 5-10 VAs - of weekly assembly work, automated end to end
- Proactive - the system flags what needs attention before anyone asks
- One model - every team reads the same governed numbers, traceable to source
Where people stay in charge
- Operators own the rules: what counts as low stock, thin margin, or drift.
- Ambiguous signals are held for a person, never auto-actioned.
- Strategy stays with the brand directors. The system just clears their fog.