Using AI Purchase Suggestions To Restock Smarter

Most sellers restock the same way: open a spreadsheet, look at last month's sales, eyeball a number, send the PO. It feels rigorous. It's actually one of the most expensive habits in e-commerce.

The reason is that "last month's sales" is the wrong signal. Demand has already shifted by the time it shows up in your sales numbers — and by the time you act on it, you're either out of stock for the bestseller or sitting on six months of inventory for the SKU that's quietly dying.

What AI purchase suggestions actually do

Forget the marketing speak. The job of an AI restock model is mundane: predict how many units of each SKU you'll sell in the next 14 to 30 days, and tell you what to order so you don't run out and don't overstock.

Good models look at three layers of signal:

• Velocity. How fast each SKU is selling, smoothed over rolling windows so a one-day spike doesn't lie to you.
• Trend direction. Is velocity accelerating or decelerating? A SKU selling 100/week and growing 15% week-over-week is a different beast from one selling 100/week and shrinking 10%.
• Cross-channel demand. If TikTok Shop suddenly spikes for a SKU, that signal usually predicts a Shopee spike 7 to 14 days later. The model picks this up; humans usually don't.

Where humans get it wrong

We've watched dozens of operations restock manually. The mistakes are remarkably consistent.

Mistake 1 — Over-indexing on the bestseller

Your top SKU is your top SKU. You over-order it because you're scared to run out. Then it does run out anyway because you ordered for last month's velocity, not next month's. Meanwhile, SKU #4 has been quietly compounding 8% week-over-week and you never noticed.

Mistake 2 — Ignoring the long tail until it dies

SKUs in positions 50-200 of your catalog get checked once a quarter. By then, half are out of stock and silently penalised by marketplace algorithms. The other half are dead and tying up cash.

Mistake 3 — Treating every channel the same

You restock based on total velocity. But TikTok shoppers want different sizes. Lazada shoppers want different bundles. If you don't break velocity down by channel, you'll oversupply one and undersupply another.

The bestseller you're most worried about is rarely the one that loses you the most money. It's the unloved SKU at position 47 that quietly stocks out for two weeks.

How Archonary's AI handles it

Our intelligent purchase suggestion runs against your full sales history across every connected store. For each SKU it produces:

• Predicted demand for the next 14 days, broken down by marketplace.
• Recommended reorder quantity given your supplier lead time.
• A confidence score — high for established SKUs, lower for new ones (so you know when to override).
• An anomaly flag when something is moving differently from its trend.

The real reason this matters

Stockouts and overstock both cost money, but they cost it differently. Stockouts cost you ranking — and ranking is compounding. A SKU that ranks #3 for its keyword and stocks out for a week often comes back at #18. Getting back to #3 takes another two months of strong sales.

Overstock costs you cash. The cash you sunk into 200 units of a SKU that's now selling 5 a week could have funded the next two trending products. The cost isn't the inventory itself — it's the opportunity cost of every PO you couldn't make.

AI purchase suggestions are not magic. They're a way of looking at every SKU every day, which no human team has time to do manually. That's the whole game.

How to start using it well

1. Connect at least two months of sales history. The model needs a baseline.
2. Set your supplier lead time per SKU. The model can't recommend reorder timing without it.
3. Review suggestions weekly, not daily. Daily is noise; weekly is signal.
4. Track override rate. If you're overriding more than 30% of suggestions, the model needs more data — or your supplier lead times are wrong.

Three months in, your stockout rate should drop by half. Six months in, your inventory cash tied up in dead SKUs should drop by a third. That's the realistic outcome — not "AI runs my business," but "AI catches the things I would have missed."