Turning daily retailer sell-through into ranked, shippable restock actions for field reps and planners
Consumer packaged goods · Just-in-time replenishment
Thread AI Editorial
September 28, 2026
Sales and supply planning at a $10B+ beverage manufacturer running direct-store-delivery
What it delivers
- $Millions in annual recovery. Lost sell-through recaptured by catching shelves before they empty
- A morning of spreadsheet work in one run. Feed reconciliation, risk scoring, forecasting, and coverage checks collapse into a single scheduled job.
- Reps receive hyper-personalized stocking instructions. Each territory gets one prioritized digest: top three SKUs per store, ranked by predicted stockout date.
- Planners only see what needs them. Coverable restocks route automatically; only uncovered shortages reach a human.
The problem
Every morning, each retail account drops a sell-through export in its own format, using its own store and SKU codes. Someone reconciles them against internal ids, estimates which stores are running thin, guesses how an upcoming promo or holiday will move demand, then cross-checks the warehouse to see what can actually ship. Only then do field reps and supply planning find out what to do.
That's a morning of spreadsheet work spread across sales, demand planning, and supply, repeated daily. By the time it lands, some shelves are already empty and the sale is gone.
The harder part is that any one of those steps in isolation produces a bad answer. Demand without coverage promises restocks the distribution center can't ship. Coverage without a forward event calendar misses the promo that's about to hit. And a reorder point alone catches the shelf that's already empty rather than the one about to be.
The approach
A Lemma Worker runs the whole sequence on a daily schedule. It ingests the retailer feeds and master sheets in parallel, normalizes three formats into one row shape, and maps every account's store and SKU codes to internal ids with territory, rep, and serving distribution center attached, so every downstream join is correct without a manual lookup.
It then scores each store-SKU against five years of sell-through history, a recent-versus-prior demand trend, and a forward event calendar, classifying it as out of stock now, at risk, watch, or healthy. For the actionable ones, a parallel AI subflow searches live web sources for real demand drivers near that store, a heat wave or a local event, and lifts the forecast only when it finds evidence. It never drops below the historical baseline.
A second subflow reconciles demand against ERP on-hand, blocked stock, committed sales orders, and inbound PO timing, allocating a shared distribution center pool by urgency, so a store can be short even when the distribution center isn't empty. Every restock gets tagged covered, covered by PO, or short.
Fully coverable work auto-emails each rep a ranked restock digest with no human involved. Only uncovered shortages escalate to a planner, in one review screen where each line already carries the reason it's short and a suggested action. Approving a line fires the purchase-order request. One human decision becomes an actioned order.
Capabilities
1
Messy multi-source ingestion into one keyed dataset.
Three retailer formats and two dimension sheets fused, normalized, and translated to internal ids in a single step, with territory, rep, and DC attached so history, events, and ERP stock all join on the same keys.
2
AI drafts the forecast, deterministic rules act on it.
The per-store demand refinement is model-generated from live web evidence and floored at the historical baseline. Coverage, allocation, ranking, and routing are all rules-driven, which is what makes the output defensible when someone asks where a number came from.
3
Tiered human-in-the-loop.
Safe work executes automatically; only the exceptions consume attention. The escalation arrives as a decision screen rather than an alert, with the shortfall, the cause, and the recommended action already computed.