Skip to content

Retail expertise

The eight retail business areas we know inside out.

Most AI teams learn your business on your budget. Our team has already spent 15+ years inside it — at Kohl's, at Sears, and running our own retail platform, HappySellers. Below is what that time bought.

01 / 08

Inventory & replenishment

Reorder points, safety stock, lead-time variance, multi-warehouse allocation.

The problem, in operator language

The bestseller stocks out again in the second week of a promo while $180K of last season's inventory sits in the wrong warehouse. Reorder quantities come from a spreadsheet built on last year's numbers. Nobody trusts safety stock levels because they were guessed once and never revisited.

What we know about it

  • Safety stock is not a single number — it changes with lead-time variance, demand variance and target service level. Most teams pick one and never revisit it.

  • Lead times are almost always understated. Real lead time = supplier promise + variance + inbound + putaway. The last three are where the misses live.

  • Multi-warehouse allocation gets wrong most often at the reorder step, not the transfer step. Reorder the wrong quantity to the wrong DC and no amount of rebalancing fixes it downstream.

What we've built for it

Reorder-point and safety-stock models per SKU per warehouse, tuned to your target service level and priced against holding cost. Multi-warehouse allocation that respects supplier MOQs and inbound cadence. Weekly reorder recommendations that go to whoever cuts POs.

02 / 08

Demand & seasonality

New-product cold start, promotional lift, size and colour curves.

Ships as part of

Demand Forecasting →

The problem, in operator language

Forecasts miss because seasonality is treated as one curve when apparel has three (weather, holiday, back-to-school) and each SKU sits on a different mix. New products get a copy-paste forecast from the closest predecessor and stock out or overbuy on day 12. Promotional lift is estimated by ops from memory.

What we know about it

  • New-product cold start is a search problem, not a forecasting problem. The right predecessor is not the visually similar SKU — it is the one with the same demand shape in the same channel mix.

  • Size and colour curves are non-stationary. The curve that worked last spring will be wrong this spring, and a curve fit on aggregate demand will be biased toward the top three sizes.

  • Promotional lift decays. A 10% price drop lifts unit sales 30% in week one, 12% in week two, and can go negative in week four if you cannibalised the next full-price cycle.

What we've built for it

Hierarchical demand forecasts (channel × category × SKU × size × colour), cold-start via a taught similarity engine, and promotional lift models with decay. Backtested on your own data so you see the honest error before we ship.

03 / 08

Pricing & margin

Elasticity, markdown timing, true per-SKU contribution.

The problem, in operator language

You discount to move stock, not because you decided to. Nobody knows true contribution margin per SKU after ads, fees, shipping and returns — so the SKUs that get promoted are the loudest, not the highest margin. Markdowns start too late and go too deep because there was no timing model.

What we know about it

  • Elasticity is not a constant. It shifts by channel, by season, by inventory position. A single elasticity per SKU is a false economy — it will lose more money than it saves.

  • Markdown timing beats markdown depth. Ten days earlier at 15% off usually recovers more margin than three weeks later at 30% off — but only if the demand curve is honest.

  • True contribution margin includes: unit cost, blended CAC, marketplace fees, payment fees, pick/pack, shipping, expected returns, and reverse logistics. Miss any two of those and the ranking changes.

What we've built for it

Per-SKU contribution margin at true unit economics, an elasticity model that respects channel and inventory state, and markdown timing recommendations that stage discounts to protect margin.

04 / 08

Customer behaviour

Segmentation, repeat purchase, LTV, churn signals.

The problem, in operator language

CAC is climbing and attribution nobody trusts. Customers buy once and disappear because there is no repeat-purchase signal and no segmentation past first-order size. The email list is treated as one audience with a discount, and the discount is what conditioned the buyer to wait.

What we know about it

  • First-order behaviour is the strongest predictor of second-order behaviour — but only when broken down by acquisition source, first-SKU category and first-order gap.

  • LTV bands beat LTV averages. A tenth-percentile customer and a ninetieth-percentile customer are almost different businesses; treating them the same optimises for neither.

  • Churn signals show up in engagement long before they show up in orders. Email open cadence, site visits, category browsing — these move first, orders move second.

What we've built for it

Behavioural segmentation with LTV bands and repeat-probability, a churn signal that fires early enough to intervene, and flow-ready outputs that plug straight into Klaviyo or a comparable ESP.

05 / 08

Merchandising & recommendations

Basket analysis, cross-sell, catalogue ranking.

The problem, in operator language

The homepage merchandises what the merchandiser noticed last week, category pages sort by rules nobody remembers writing, and product-detail-page recommendations are 'people also bought' — which is a proxy for popularity, not relevance. Cross-sell is left on the table because nobody has time to model it.

What we know about it

  • 'People also bought' overweights hero SKUs and starves the long tail. A recommendation model that ranks by lift, not raw co-occurrence, exposes a much bigger catalogue.

  • Recency matters for basket analysis but is usually ignored. A basket from this quarter carries more signal about current intent than one from a year ago.

  • Catalogue ranking is the single highest-leverage merchandising decision. The order SKUs appear in a category page decides 60–70% of what gets sold.

What we've built for it

Recommendation engines architected the way our team built them for 20M+ Kohl's shoppers, sized down for a Shopify catalogue. Basket analysis and cross-sell for PDP and cart. Catalogue ranking that respects business rules and business objectives at the same time.

06 / 08

Returns

Return risk at checkout, size-driven returns, cost allocation per SKU.

Ships as part of

Demand Forecasting →

The problem, in operator language

Returns eat 8–20% of revenue in apparel and 20%+ in footwear, and the P&L only shows the aggregate. Nobody knows which SKUs are structurally lossy after returns are counted, which customers are serial returners, or which promotions triggered the last spike in returns.

What we know about it

  • Returns are largely predictable at checkout. Size-fit ambiguity, category, first-time-buyer signal, and past return behaviour together explain most of the variance.

  • The cost of a return is not just reverse shipping. It is inspection, restocking, obsolescence, and — for apparel — the shrink from resell-as-B-grade.

  • Discount-driven cohorts return at higher rates. Optimising promo depth without modelling return-adjusted margin systematically overpromotes.

What we've built for it

Return-risk scoring at checkout, size-driven return prediction with fit-clue models, and per-SKU cost allocation that shows honest margin after returns. Feeds pricing, promo and merchandising decisions.

07 / 08

Marketplace & wholesale ops

Settlements, deductions, chargebacks, sell-through after shipment.

The problem, in operator language

Amazon and Walmart settlements arrive as thousand-line files that ops reconcile by hand. Deductions and chargebacks get eaten because nobody has time to dispute inside the window. Wholesale sell-through data lives with the buyer, weeks late, and nobody knows real velocity until reorder time.

What we know about it

  • Marketplace settlement lines have a repeating grammar. Once you learn it, 90%+ of exceptions can be matched automatically and only the truly odd cases need a human.

  • Chargebacks and deductions are often recoverable, but only inside a short dispute window. Missing the window is where the money actually goes.

  • Wholesale sell-through, when reported at all, is at the account level. Getting it back to SKU × door × week takes an ELT the buyer never built.

What we've built for it

Overnight reconciliation across Shopify, Amazon, Walmart, 3PLs and payment processors — matched by pattern, exceptions to a human. Chargeback and deduction detection with dispute-window alerts. Wholesale sell-through pipelines that turn buyer reports into weekly SKU × door velocity.

08 / 08

Retail data engineering

Shopify, Amazon, ERP, POS, 3PL, ad platforms into one trusted model.

Ships as part of

Demand Forecasting →

The problem, in operator language

Shopify says one number, Amazon another, the 3PL a third, the books a fourth. Weekly reporting eats two days of someone's week. Every new tool added a dashboard but no answer. There is no single version of the truth to build a forecast on top of, so every model built downstream inherits four different definitions of 'sale'.

What we know about it

  • The definition problem is bigger than the pipeline problem. Two teams call a return an 'order' and a 'negative order' respectively, and neither is wrong — until they need to reconcile.

  • Ad-platform data is the last mile people underestimate. Meta and Google attribution windows do not agree with anyone else, and blending them naively double-counts.

  • A trusted model is worth more than a fast one. Latency matters, but only after correctness. Most retail decisions run on a weekly cadence and can afford a slow pipeline that is right.

What we've built for it

A retail data pipeline pattern that ingests Shopify, Amazon, Walmart, ERP, POS, 3PL, Klaviyo, Meta and Google Ads into one model with a shared definition layer. Weekly refresh minimum, hourly where a decision demands it. Documented, auditable, transferable.

Which one moves your number first?

Book a 30-minute call.

We'll rank the eight areas by ROI for your business right now, and tell you honestly which one to ship first. If none of them pay back for you today, we'll say so on the call.

The Retail AI Implementation Weekly

Practical AI implementation for e-commerce operators. No hype.