Automated PO generation. The system drafts the reorder before you notice the gap.

The founder brief named it directly: warehouse leads were guessing at restock priorities. With 16,000 SKUs across multiple warehouses and seasonal demand, no operator can track which items are about to run out, in which location, and how long each supplier needs to deliver.

Reorders went out late, in the wrong quantity, or from a slower supplier than necessary, and stockouts in the highest-velocity SKUs cost the most. Before this feature, buyers pulled inventory reports into spreadsheets, cross-referenced supplier lead times from memory or email, and built POs line by line. A single buyer's weekly reorder pass could take most of a morning; by then, fast-moving SKUs had already dipped below safety stock.

• Detect. The system monitors stock levels against reorder thresholds across every warehouse continuously. Not a fixed reorder point alone. Thresholds are driven by the 7-day demand-forecasting model: projected depletion date and safety stock together determine when a SKU enters the "about to run out" queue.
• Suggest quantity. Reorder quantity combines forecasted demand over the supplier's lead time, current on-hand and in-transit stock, and the supplier's minimum order quantity. The goal is enough to cover demand until the next delivery window without over-ordering slow movers.
• Suggest supplier. When a SKU has multiple approved suppliers, the pick weighs lead time first (can they deliver before stockout?), then unit cost, then historical on-time performance. The buyer sees why this supplier beat the alternatives.
• Draft the PO. Accepted suggestions assemble into a complete purchase order, line items, quantities, supplier, delivery warehouse, and expected dates, ready for one-click approval in the Purchasing module.

Restock suggestions live in a dedicated queue within the Purchasing & PO module. Not buried in notifications. Each item is an AI cell stacked above a collapsed PO preview: line items, supplier header, delivery warehouse, and total value visible without expanding.

• One-click approve. Primary action on every suggestion. The PO moves to "Submitted" immediately; the suggestion clears from the queue.
• Edit before approve. Expanding a suggestion opens the full PO form with every field pre-filled. Quantity, supplier, and warehouse are editable inline; changes are logged as overrides.
• Split across warehouses. When the same SKU is low in two locations, the suggestion can generate separate line items per warehouse on one PO, or two POs if suppliers differ.
• Batch review. Buyers can filter the queue by warehouse, category, or urgency (days to stockout) and work through high-velocity SKUs first.

Restock queue · 12 suggestions ready, high-confidence auto-drafts flagged

Suggestion detail · "Why this?", stock, 7-day forecast, lead time, last price

Edit before approve · supplier alternates + override memory note

Dismiss with reason · structured feedback improves future suggestions

1 / 4

The first version shipped with confidence scores but no memory of overrides. Buyers corrected the same suggestions again and again, and accept-rates plateaued around 40%. Once the model remembered overrides , learning a buyer's supplier preferences and quantity adjustments, accept-rates climbed to 64%. Confidence scores are table stakes; override memory is what builds trust. (Lesson from the HikeOn case · Section 08)

Beyond the override-memory story, harder states needed explicit design:

• No supplier meets the timeline. The suggestion shows the fastest available option with a "will stock out before delivery" warning. Buyer chooses: order partial now, expedite, or accept the gap and dismiss.
• MOQ exceeds need. Quantity rounds up to supplier MOQ with a note on overstock risk. The "why this?" explainer shows the MOQ constraint so the buyer isn't surprised by the larger number.
• Volatile demand. Forecast confidence drops below threshold; the suggestion opens in edit mode with a "review recommended" badge. No one-click approve on low-confidence items, edit or dismiss only.
• Split across warehouses. Same SKU low in two locations generates one suggestion with two line items. Buyer can approve both, edit one, or split into separate POs if suppliers differ per warehouse.
• Dismissed twice. A SKU dismissed twice in 30 days leaves the queue for that buyer and escalates to the purchasing lead's view, so chronic false positives don't hide real gaps.

After override memory shipped, 64% of restock suggestions were accepted without editing, up from ~40% when every correction was forgotten by the next run. Buyers report the weekly spreadsheet pass is gone; reorders surface in the Purchasing queue as drafted POs instead of blank forms.

The accept rate is the headline metric on the card because it measures whether the draft is good enough to ship. One-click approve only works if the suggestion is right often enough that buyers trust it, override memory was the difference between a demo feature and a daily workflow.