- Market
- Saudi Arabia
- Written for
- Commercial Director, Head of CX
- Decision
- Use AI on demand, inventory and personalisation to move margin
Across a season
Where AI touches a season, from the buy to the markdown
Retail AI is not one product; it is a set of decisions strung across a season, each with its own P&L consequence. These are the points where a model changes the number, and each is judged on the money it moves.
Sizing the buy
A buyer commits to quantities months ahead, and an over-buy becomes a markdown while an under-buy becomes a stockout.
A forecast built on your own history and drivers sizes the buy closer to real demand than a rule of thumb carried from last year.
- Sales history
- Forecast
- Buying
Allocating to stores
The same product sells differently across branches, so a flat allocation strands stock in one and starves another.
Allocation follows each store's expected demand, and reallocation shifts stock between them before either extreme becomes a loss.
- Store demand
- Allocation
- Transfers
Making the offer personal
A customer sees a generic promotion that either wastes margin on someone who would have bought anyway or misses them entirely.
Personalisation targets the offer where it changes a decision, and is judged on the incremental sale rather than the click.
- Customer data
- Personalisation
- Channels
Clearing the tail
End-of-season stock is marked down late and deep, giving away margin that earlier action could have kept.
The model flags slow lines early enough to act with a shallower markdown or a transfer instead of a deep clearance.
- Sell-through
- Markdown
- Pricing
The recommendation trap
A recommendation that sells what would have sold anyway
A personalisation engine can look like a triumph and add nothing. If it recommends to a customer the thing they were already going to buy, the click-through is high, the dashboard is green, and the margin is unchanged — you paid to influence a decision that was already made. The number that matters is not whether a recommendation was clicked but whether it caused a sale that would not otherwise have happened, and those two numbers can point in opposite directions. A model optimised for clicks will happily learn to take credit for demand it did not create.
The same discipline applies to a forecast. A model that reports a lower error against last season's numbers has not helped if the error it reduced was on lines that were never a problem, while the volatile, hard-to-call items — the ones that actually drive markdowns and stockouts — are no better predicted. So we measure retail AI against the money: does the forecast reduce the markdowns and stockouts that hurt, does the allocation move stock to where it sells, does the personalisation cause incremental margin. A model that improves a metric no one was losing money on is a science project, not a retail tool.
Where the money leaks
Four ways retail AI quietly loses the margin it promised
None of these show up in a demo. They show up in the season's numbers, which is why we design against them before a model is chosen.
- 01
Forecasting the easy lines
A model can cut average error while missing exactly the volatile items that cause the markdowns, so the headline improves and the losses do not.
- 02
Personalising to existing intent
Recommending what a customer would have bought anyway inflates click-through and adds no incremental margin, paying for demand that already existed.
- 03
Ignoring what is on the shelf
A forecast disconnected from live stock recommends selling what is not there and reordering what is overstocked, undoing its own gains.
- 04
A model the floor does not trust
When a buyer or store manager cannot see why the model says what it does, they override it, and a recommendation nobody follows moves no margin at all.
Our role
What we build and what stays a commercial call
- 01
Measured on your P&L
We define the margin a use case is meant to move before we build it, and test the model against your own history so the benefit is proven in money, not in an error metric.
- 02
Connected to real stock
We tie forecasting and allocation to your live inventory and POS, so a recommendation reflects what is actually on the shelf rather than a number in a separate model.
- 03
Trusted by the people who act on it
We make the model's reasoning visible to buyers and store teams, because a recommendation they understand is one they act on, and one they act on is the only kind that moves margin.
The buy, the price and the markdown are commercial decisions that stay with your buyers and merchandisers; we build the forecasting, allocation and personalisation that inform them and prove their worth on the season's margin rather than on a demo.
Common questions
What buyers ask
How do we know a forecasting model actually beats what our planners do now?+
By backtesting it against your own history and comparing it to your planners on the decisions that cost money, not on average error. We run the model over past seasons as if it had been making the calls, then measure the markdowns and stockouts it would have caused against the ones you actually had. A model that lowers average error but does no better on the volatile lines that drive your losses has not earned its place. The honest comparison is on margin — would the buy, the allocation and the markdowns have been better — and if the model cannot beat your planners on that, we would rather say so than deploy it.
Does personalisation just recommend what customers would buy anyway?+
A badly measured one does exactly that, which is why we evaluate personalisation on incremental margin rather than click-through. If a recommendation lands on a customer who was already going to buy the item, it inflates the click rate and adds nothing to the P&L — you have paid to be credited for a sale you already had. The right measure is whether the recommendation caused a purchase that would not otherwise have happened, tested by comparing against a held-back group who did not receive it. We build and judge personalisation on that lift, so it targets the decisions it can actually change and does not flatter itself with demand it did not create.
Can it work with our existing POS and inventory systems?+
It has to, because a forecast disconnected from your live stock does more harm than good. A model that recommends selling what is not on the shelf, or reordering what is already overstocked, undoes its own gains and loses the trust of the people meant to act on it. We connect the model to your POS and inventory so its recommendations reflect what is actually available, and we work with the systems you have rather than requiring you to replace them first. Where the data is messy — and it usually is — cleaning and joining it is part of the work, not a prerequisite you have to finish before we start.
In this cluster
Next step
Map it to your operation
A read on one use case — forecasting, allocation or personalisation — that names the margin it is meant to move and how that will be measured against what you do now.
Map it to your operation
What happens next
- 1We reply within one business day.
- 2A 30-minute scoping call with a senior consultant.
- 3A written proposal with scope, team, and timeline.
Or reach us directly

