Retail margins ride on inventory turns. A product out of stock for three days costs more than three days of revenue — it costs the customer's next decision.

Tianhong's AI program started exactly there, at the most painful link in that chain.

Three workflows that changed

Merchandising selection. Buyers used to search, compare and read trends across several platforms. Now an agent combines procurement group chats, store membership data and external social heat to propose a candidate list. Buyers judge the list rather than assemble it.

Stockout attribution. When an item ran out, staff manually traced order, inventory and delivery records to find the cause — often two to three days of digging. The agent now aggregates system data and pushes alerts to stores, compressing attribution to the same day.

Complaint analysis. Customer feedback was scattered across enterprise chat groups, impossible to read in full. The agent consolidates and analyzes it, handling roughly 100,000 messages per month.

Combined, Tianhong expects to save about 1,800 work hours a month.

The hard part was never the model

All three workflows share one precondition: data has to be reachable. Selection needs membership data. Attribution needs orders, inventory and delivery. Complaint analysis needs the chat platform.

Plenty of companies stall here. The model call works, the integration does not. The integration works, the fields do not match. The fields match, permissions block access.

Tianhong's approach was to put the agent inside WeCom, the app employees already use all day, then bridge internal systems through connectors. Nobody had to learn a new tool, and the workflow was not cut in half.

Two design choices worth copying

First, the AI does not decide. Selection produces candidates; attribution produces alerts and probable causes. Humans still commit. In retail, where a wrong decision is expensive, that boundary is not optional.

Second, the benefit is measured in hours. 1,800 hours is auditable, which makes it far easier to defend the budget in year two. "Improved efficiency" does not survive a planning cycle.

A replication path

If you run retail or a chain operation, the sequence looks like this:

1. Pick a link where data is already digital and outcomes are unambiguous — stockout attribution is a good first target.
2. Align fields across systems. This step eats the most calendar time; do not underestimate it.
3. Start with recommendations, not automatic writes to inventory.
4. Accept the project on hours saved or days shaved.

The competitive gap in retail AI is data plumbing and process granularity. Model capability will converge. How deep your systems connect and how finely you decompose the workflow will not.