Inventory Recommendations Should Improve Over Weeks

What happens after you approve

You approve a markdown on a batch of yogurt Tuesday morning. Then the week happens. Deliveries, a rainy weekend, a promo down the street. By Friday the batch is either gone or sitting in the dumpster, and nobody writes down which.

That missing note is the whole problem. A recommendation, whether it comes from software, a consultant, or your own gut, is a guess. It only turns into knowledge once you compare it with what actually happened. Did the discounted units sell through? Did the reorder you trusted sit on the shelf for three weeks? Without that comparison, next Tuesday's call is built on exactly the same assumptions as this one.

So when a tool promises perfect inventory recommendations from the first day, be suspicious. The honest version says something different: it'll be roughly right at first, and it gets better as it learns how your store behaves.

Where a first-week call comes from

On day one, any recommendation engine knows only what's in your records: stock on hand, batches, cost, expiry dates, and whatever sales history you exported. From that it can work out a sell-through rate and project forward. That's useful. It's also blind in a few predictable places.

• Markdown response. It doesn't know how hard your customers react to a discount. A shopper at a busy suburban grocery and a shopper at a corner store near an office park behave differently when they see a yellow sticker.
• Category quirks. Bread might move fast once it's marked down while specialty cheese barely moves at all. The records don't say that until you've tried it.
• Supplier reality. The lead time printed on a vendor sheet and the lead time you actually get aren't always the same. The gap shows up in your purchase orders, not in the catalog.
• Your approval habits. If you always shave a recommended discount down a bit, the tool needs to know, or it'll keep overestimating what the markdown will do.

None of these can be known in advance. They can only be observed. That's why "learns over time" is the right promise and "instantly optimises" is the wrong one. We go into the wider gap between hype and reality in AI inventory management: real vs hype.

The feedback loop, step by step

A real inventory agent feedback loop has four parts, and if any one of them is missing the learning stops.

• The call, with its arithmetic. The recommendation has to say what it expects: how many units won't clear at the current pace, what discount it suggests, and how much it expects to recover. If there's no stated expectation, there's nothing to check later.
• The decision. You approve it, adjust it, or reject it. Each of those is information. A rejected call tells the tool something about your judgment, too.
• The outcome. What actually happened to that batch or that reorder: sold, wasted, still sitting there.
• The adjustment. Next time a similar situation comes up (same category, same kind of discount, same supplier), the projection shifts toward what really happened.

Step three is the one that tends to break. Outcomes aren't announced. They show up quietly in the next stock export as units that dropped or didn't. A tool that reads your records every day can pick that up without anyone filling in a form. A tool that only looks once, or only when you open it, can't.

What counts as an outcome

Learning from markdown results means being specific about what "worked" means. Selling the batch isn't enough on its own, because a deep enough discount sells anything.

Useful outcome questions:

• Did the marked-down units clear before the date, and how many were left over?
• Did the money recovered come close to what the recommendation projected?
• Did the markdown eat into full-price sales of a fresher batch of the same item?
• For a reorder: did the new stock sell through before the next delivery, or did it pile up behind older batches?

A tool that only counts "approved and sold" will learn to recommend steeper and steeper discounts. A tool that compares projected recovery with actual recovery learns to find the smallest discount that does the job. If you want to see what a well-built markdown call looks like before any of this learning kicks in, see dynamic markdown pricing for near-expiry products.

A worked example over four weeks

Worked example, week one: a store has 14 units of a yogurt SKU with five days left, selling at 0.4 units a day. At that pace about 2 units clear and roughly 12 won't. The tool assumes a discount will roughly double the sell rate and recommends a markdown. You approve.

Worked example, week two: the next export shows 9 units sold and 5 written off. The discount helped, but less than projected. The lift for this category at this store was smaller than the starting assumption, so the projection gets nudged down.

Worked example, week three: a similar batch comes up. Now the tool flags it a day or two earlier, because it's learned that a short markdown window doesn't give enough time to clear the units here. Starting earlier at the same discount gives the batch more selling days.

Worked example, week four: the earlier call clears all but one unit. Nothing about the store changed. What changed is that the tool stopped assuming and started using what it had seen.

That's the shape you want: early calls that are reasonable but imperfect, then a visible drift toward your store's real behaviour. Our guide on checking the arithmetic in a markdown recommendation shows how to read each call so you can watch that drift yourself.

Questions to ask any tool

Before trusting a system with recommendations, ask:

• Does every call show its arithmetic? Units, pace, discount, expected recovery. If it can't show its working, you can't tell whether it's improving.
• Does it record what I approved and what I changed? Your edits are training data.
• Where does it get outcomes from? It should read them from your own stock records over the following days, not from a survey you'll never fill in.
• Does it go quiet when it isn't sure? A tool that's still learning a category should hold back on calls there rather than guess loudly. See inventory alert fatigue for why fewer, better calls matter.
• Does reorder sizing use my real lead times? Lead time should come from your own purchase orders. We cover the method in supplier lead time from purchase orders.

Running the loop by hand

You can do this without software. Keep a simple log: date, item, batch, what you decided, what you expected, and then (the step everyone skips) what happened a week later. Review it every Monday. After a month you'll start to see patterns, like which categories respond to discounts and which suppliers run late.

What a log can't do easily is stay consistent. It depends on someone remembering to close out every line, and it gets harder with every SKU you add. It also won't recalculate next week's projections for you. You'll carry the lessons in your head, which works until you're off for a few days.

Where ShelfSense fits

ShelfSense by ShelfLifePro is a daily inventory agent built around this loop. It works alongside the system you already run. Once a day it reads your stock, batch, cost and expiry records and surfaces what to mark down, clear or reorder, ranked by money at risk, with the arithmetic shown. It recommends and you approve. It never changes a price or places an order on its own, and it learns over time from which accepted calls sold and which got wasted. You can start with a CSV export of your stock and expiry dates. If you're on Tally, Shopify or WooCommerce, you can export to CSV today, and a direct connection is available on request.

To see how a call gets built and then corrected, read how ShelfSense thinks. The Watch tier is free forever at 30 SKUs — a daily scan and a health report.