AI for physical retail: what actually changes in a store

Talk about AI for physical retail and the conversation drifts to ecommerce within two minutes: recommenders, site search, abandoned carts. The store itself gets left out, which is odd, because that is where most retail revenue still happens and where every mistake costs physical money: square metres occupied by stock that doesn't move, empty shelves on a Saturday afternoon, staff spread across the wrong hours.

This guide covers what a chain of 5 to 200 stores — fashion, grocery, hardware, beauty, electronics — can do today with the systems it already has: POS, ERP, cameras and, if you're lucky, a loyalty programme. It isn't about checkout-free stores or shelves that weigh their own stock. It's about footfall, assortment, replenishment and shrink, which is where data changes a decision this week.

A warning up front: if nobody in your chain can answer within two minutes how much each store sold yesterday by product family and at what time, the problem isn't AI yet. It's scattered data between POS and ERP. It gets fixed sooner, costs less, and everything else depends on it.

What can AI do in a store that the POS doesn't already do?

The POS tells you what sold. It doesn't tell you what didn't sell, or why. That gap is most of the value advanced analytics brings in store:

• Lost sales. The POS records the 40 pairs of trousers sold, not the 12 customers who looked for a size 42 and left. Crossing entry footfall with tickets (conversion by time slot) and stock by size starts to show that picture.
• The cause of a gap. A product with flat sales may be dead, or it may have gone three weeks without being restocked on the shelf. Those are opposite decisions and the sales report shows them identically.
• Patterns you can't see by walking the floor. Which products get bought together only in shopping-centre stores, which references pull the whole basket despite low units, which store behaves nothing like its peers.
• Anticipation. Forecasting footfall by store, day and time slot lets you set shifts and orders a week ahead instead of reacting on Friday.

None of this needs an exotic model. It needs sales, stock and footfall data in one place, clean, with a single definition of "store", "product" and "day". That groundwork is what data governance covers, and it's the part the vendor demo skips.

Which systems do you have to touch: POS, cameras, ERP?

Before promising anything, look at where each piece of data comes from and what it costs to extract. In a mid-sized chain the map usually looks like this:

| Source | What it gives you | Usual problem | |---|---|---| | POS | Ticket, line item, time, payment method, operator | Each store on its own version; promotions badly coded | | ERP / stock management | Theoretical stock by store and SKU, orders, receipts | Theoretical ≠ real stock; 5-15% drift without frequent counts | | People counter / entrance camera | Footfall by time slot, conversion rate | Counts staff and couriers if badly configured | | Shop-floor cameras | Heatmaps, queues, incidents | Personal data: needs a legal basis and on-device analysis | | Loyalty / CRM | Identified customer, frequency, basket | Covers only 20-50% of tickets | | Planogram / visual merchandising | What should be on each shelf | Lives in a PDF or in the area manager's head |

The practical rule: start with the sources that already produce structured data (POS and ERP) and add footfall next. Shop-floor cameras are the last layer, not the first — not for technical reasons but because of installation cost and compliance.

Integrating POS and ERP isn't glamorous, but it's 60% of the project. If you're going to do it, read the patterns and failure modes in ERP and CRM integration first: they transfer almost unchanged.

Why count footfall if you already know what you sell?

Measuring footfall changes the unit of analysis: you stop comparing sales across stores and start comparing conversion. A store selling half as much as another may be doing better if it gets a third of the visits.

With footfall in 30-minute slots and POS tickets you get three actionable things:

1. Staff scheduling. If 38% of visits come in between 18:00 and 20:00 and you have one person on the till then, the conversion problem isn't the assortment. This is solved with a footfall forecast by store and day — a simple model (history, day of week, holidays, weather, retail calendar) and a verifiable one: every Monday you compare forecast against actual. 2. Promotion effect. A campaign that lifts sales 20% but footfall 35% isn't working: it's pulling in people who don't buy. 3. Per-store anomaly detection. A 30% conversion drop in one store for four consecutive days is almost always a stockout or a staffing problem — and you only catch it if an alert fires automatically.

Counting people isn't the project. The project is that somebody gets a short list of anomalous stores at 9:00 on Monday and has to explain what happened.

How do you tune assortment store by store without losing your mind?

Assortment is where a chain with heterogeneous stores leaves the most money. Two classic mistakes: one assortment for everyone ("whatever works in the flagship") or assortment at each manager's discretion (ungovernable and incomparable).

The middle path that works:

• Cluster stores by behaviour, not by floor area. A clustering model over sales mix, seasonality, basket type and customer profile usually yields 4 to 8 useful groups. Same logic as customer segmentation, applied to stores.
• Define a common core and a flexible tail. Typically 60-70% of references shared chain-wide and 30-40% decided per cluster. That keeps supplier negotiation and buying power intact.
• Measure cannibalisation before expanding. Adding 200 new references to a 300 m² store doesn't add sales: it redistributes space. The right question is what comes out, not what goes in.
• Put a review date on it. Cluster assortment without a quarterly review ages in a single season.

For this to be more than a spreadsheet exercise you need rotation by SKU and store, real margin (after discounts and shrink) and space occupied. If margin per reference is an estimate, the model will optimise the estimate.

How much of replenishment can be automated?

Replenishment in physical retail has two levels and it pays not to mix them.

From central warehouse to store. Here AI genuinely contributes: demand forecasting by store and SKU, safety stock calculated from variability instead of a fixed rule, and an automatic order proposal that a manager approves. It's the case with the most measurable ROI, and the one covered in inventory optimisation and demand forecasting.

From the stockroom to the shelf. Here the limit is that nobody knows what's actually on the shelf. The ERP says "6 left" but those 6 may be in the back, in a fitting room or stolen. Options, cheapest first:

1. Zero-sales alert: a reference with positive theoretical stock and zero sales for X days in a store whose cluster peers are selling it. Cost: no hardware. Reasonable hit rate. 2. Targeted cycle counting: instead of counting everything twice a year, the system prioritises the 50 references most likely to be out of sync each week. 3. RFID or computer vision on shelf: high accuracy, high cost. Only worth it in high-ticket or high-shrink categories. If you're considering cameras, read computer vision in business first.

Order matters: the zero-sales alert costs a few weeks of data work and usually recovers more sales than an RFID pilot that takes a year to roll out.

And shrink? Where's the limit?

Unknown shrink runs around 1-1.5% of sales in European retail, and a big share of it isn't external theft: it's administrative error, badly recorded waste and receipt discrepancies. Attack it in that order, because you already have the data.

What you can do with your own data and no camera:

• Till exception analytics: voids, returns without receipt, manual discounts and out-of-pattern drawer openings, by operator and by store. This is classic anomaly detection and the best cost-to-result ratio on the list. Same logic as fraud detection.
• Receipt reconciliation: compare delivery note, order and actual receipt to find the supplier who systematically delivers short.
• Waste traceability: if each store records waste by hand with free-text categories, no analysis is possible. Normalise first, analyse second.

What needs care: video analytics on customers or staff processes personal data. You need a legal basis, clear notice, an impact assessment where processing is large-scale, minimisation (analyse on-device and avoid storing images wherever possible) and a short, written retention period. Facial recognition to identify specific individuals is very-high-risk territory and in practice isn't a reasonable route for a retail chain. Before signing with a video analytics vendor, run the case through GDPR and AI.

Where should a store chain start?

A realistic order by effort-to-result:

| Order | Project | Data it needs | Sign it's your turn | |---|---|---|---| | 1 | Unified sales and stock by store/day/SKU | POS + ERP | Every report takes days and each team has its own number | | 2 | Stockout and zero-sales alerts | The above | Customers asking for what isn't there; empty shelves | | 3 | Till exceptions | POS log | Unknown shrink above 1% | | 4 | Footfall and conversion by time slot | Entrance counter | Doubts about shifts and opening hours | | 5 | Demand forecasting and order proposal | 18-24 months of clean history | Excess stock and stockouts at the same time | | 6 | Assortment by store cluster | Margin and space per SKU | Very different stores on the same assortment | | 7 | Computer vision on the shop floor | Installation + legal basis | Only once 1-6 are solved |

None of these starts by buying a tool. They all start by deciding which decision will change and who makes it. If there's no named person who will act on the system's output, the system is surplus — the same point we make in change management.

Frequently asked questions

How much history do you need to forecast store demand?

Between 18 and 24 clean months, to capture two full seasonal cycles. With 12 months you can start in fast-moving, low-seasonality categories, accepting more error. The critical factor isn't volume but consistency: product recoding or changes to the store estate break the series.

Can you count footfall without breaching GDPR?

Yes, if the counter doesn't identify people. Sensors that count silhouettes or line crossings and only store an aggregate number per time slot don't process personal data. The problem starts with cameras that store images or profile by age and gender: that is processing, and it needs a legal basis and notice to the people affected.

What if my stores are franchises?

The governance changes, not the technique. You have to agree contractually what data flows to head office, how often and in what format, and what the franchisee gets back (forecast, order proposal, comparison against its cluster). Without that trade, data arrives late and dirty.

Is it worth it with only five stores?

For forecasting and cluster assortment, probably not: with five stores a good manager and a solid spreadsheet go a long way. For unified sales and stock, and for till exceptions, yes: those cases scale down well and cost little.

---

If you have stores, a POS and the feeling that the data is there but nobody uses it to decide, the first step isn't an AI pilot: it's knowing what you can answer today with what you already have and what's missing for the next step. That's exactly what we look at in the audit, and if you'd rather talk it through in half an hour, let's talk.