AI for CPG and FMCG Companies
The defining constraint in consumer goods is that you do not own the shelf, the price or the customer relationship, so most of your data about what actually happened is bought, delayed and incomplete.
Consumer goods companies routinely model what they shipped rather than what consumers bought, because shipments are their own data and consumption is not. That single substitution explains a great deal of the sector's forecasting disappointment.
AI for CPG and FMCG covers consumption-based demand forecasting, promotional and trade spend effectiveness, retail execution and availability monitoring, assortment and space optimisation, revenue growth management, and new product forecasting.
Shipments are not demand
What you ship reflects retailer ordering behaviour, inventory policy, promotional buy-in and forward-buying ahead of price increases. What consumers bought is a different series, and the gap between them is where most forecasting error lives.
| Series | What it actually measures | Where it misleads |
|---|---|---|
| Shipments | Retailer ordering behaviour | Forward buying and inventory swings look like demand |
| Retailer EPOS | Consumer purchases at that retailer | Best available; coverage and latency vary by account |
| Syndicated panel | Estimated category consumption | Sampling error, and it lags |
| Promotional lift | Sales during promotion | Includes pantry loading and post-promotion dips |
| Distribution data | Where you are listed | Being listed is not being on the shelf |
Model consumption, then model the ordering behaviour on top
The structure that works is to forecast consumer demand from EPOS and panel data, then model retailer ordering as a separate layer with its own inventory dynamics, promotional buy-in and forward-buy behaviour. Collapsing the two into a single shipment forecast is faster and it embeds every retailer inventory swing into your demand signal permanently. This is the single most valuable structural change we make in CPG engagements.
Trade spend, the largest and least measured line
Trade spend is frequently the second largest line on a consumer goods P&L and one of the least rigorously evaluated. The reason is not indifference — it is that promotional measurement is genuinely hard, and the easy measurement overstates every promotion.
Baseline estimation is the whole game
Lift is sales minus what would have happened anyway, and that counterfactual is unobserved. A baseline estimated from surrounding weeks is contaminated by pantry loading before and the dip afterwards. Getting the baseline right matters far more than the sophistication of the lift model.
Count the dip and the cannibalisation
A promotion that sells volume by pulling forward next month's purchases and stealing from your own full-price line has moved units without creating value. Measure over a window long enough to include the post-promotion trough, and across your own portfolio rather than one SKU.
Incremental margin, not incremental volume
Deep discounts reliably move volume. Whether they move contribution is a different question and frequently has a different answer, particularly once trade funding and the base-price erosion effect are counted.
Some promotions should be stopped
In most portfolios a meaningful share of promotional activity is value-destroying and continues because it always has. Identifying those is uncomfortable, valuable, and usually the first real result of this work.
Retail execution, where the shelf is not what the plan says
- Listed is not stocked, and stocked is not on shelf. Phantom availability — inventory recorded but not on the shelf — is a persistent and expensive gap.
- Field team photos are a usable data source. Shelf images from field visits support planogram compliance, share of shelf and out-of-stock detection at reasonable cost. See object detection.
- Prioritise the visits. Field team time is fixed, so predicting which stores are most likely to have an execution problem is worth more than measuring all of them equally.
- Share of shelf against plan is the honest metric. Not whether a planogram exists, but what is actually on the fixture this week.
- Retailer data-sharing determines what is possible. Your visibility differs enormously by account, and any model has to work with that unevenness rather than assume it away.
How an engagement runs
Consumption first, ordering behaviour second — because collapsing them is the sector's standard error.
Data landscape
Which series you have per account, at what latency and coverage, and where consumption is genuinely observable.
Consumption model
Consumer demand from EPOS and panel, with promotional and seasonal effects modelled explicitly.
Ordering layer and trade spend
Retailer ordering behaviour modelled separately; promotional baselines and incremental margin estimated.
Validation against held-out periods
Including promotional periods, which is where naive models fail most visibly.
Deployment into planning
Into the demand planning and trade planning cycles, with accuracy tracked by account and by promotion type.
What you receive
A demand signal that reflects consumers, and a trade spend picture that reflects margin.
Data landscape assessment
What each series measures, its coverage and latency by account, and where it misleads.
Consumption-based demand forecast
Consumer demand separated from retailer ordering behaviour.
Retailer ordering model
Inventory dynamics, promotional buy-in and forward buying as their own layer.
Promotional baseline and lift
With pantry loading, post-promotion dip and portfolio cannibalisation counted.
Trade spend effectiveness
Incremental margin by promotion type, account and SKU — including the negative ones.
Retail execution prioritisation
Which stores to visit, based on predicted execution problems rather than rotation.
Is this the right starting point?
Worth being direct. There are situations in consumer packaged goods where custom AI work is the wrong spend, and those are listed rather than buried.
Worth doing if
- Your demand forecast is built on shipments and behaves badly around promotions.
- Trade spend is a major P&L line and its return is estimated rather than measured.
- You have retailer EPOS access for meaningful accounts and are not using it well.
- Field team visits are scheduled by rotation rather than by predicted need.
- New product forecasting is consistently and expensively wrong.
Do something else if
- You have shipments only, with no EPOS or panel access. Consumption cannot be recovered.
- The demand planning process cannot change what it orders inside the forecast horizon.
- Trade terms are set by negotiation regardless of what the analysis says.
- You want price coordination across the category. That is a competition law problem, not a model.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Why is our forecast always wrong around promotions?
Usually because it is trained on shipments. Retailers buy in ahead of a promotion and buy less afterwards, so a shipment series contains inventory swings that look like demand and get learned as seasonality. Forecast consumption from EPOS and panel data, then model retailer ordering as a separate layer — that structural change usually does more than any modelling improvement.
How much of our trade spend is actually working?
In most portfolios less than the plan assumes, and a meaningful share is value-destroying. The difficulty is the baseline: lift is sales minus what would have happened anyway, and the surrounding weeks are contaminated by pantry loading before and the dip after. Getting the baseline right matters far more than the lift model, and the first honest output is usually a list of promotions to stop.
Can we use shelf photos from our field teams?
Yes, and it is one of the better-value applications in the sector. Shelf images support planogram compliance, share of shelf and on-shelf availability detection without new hardware. The more valuable step is using the results to prioritise where the field team goes next, because their time is the fixed constraint and rotation-based scheduling wastes it.
What about new product forecasting?
Hard, and improvable. There is no history for the product itself, so the workable approach is attribute-based analogues — forecasting from comparable launches with similar attributes, price positioning, distribution plan and category dynamics. Be honest about the uncertainty range rather than reporting a point estimate, because a confident single number on a launch forecast is misleading by construction.
Can AI help us set prices across the category?
For your own pricing and revenue growth management, yes — elasticity estimation, price pack architecture and promotional depth are all tractable. Anything that coordinates pricing with competitors, or that uses a shared algorithm to align market prices, is a competition law question rather than an engineering one, and we would want your legal team to answer it before we build anything.
Related verticals
Organisations in consumer packaged goods usually share data, buyers or regulators with these. All fourteen are listed on the Retail & E-commerce page.
AI for Grocery Retailers
Fresh forecasting, waste reduction, markdown and substitution — where a forecast error becomes food waste.
Read more →AI for Brick-and-Mortar Retail
Store-level forecasting, allocation, scheduling and shrink — with in-store vision assessed honestly.
Read more →AI for Beauty and Cosmetics Brands
Shade matching, ingredient data, review integrity and trend-driven forecasting — with claims treated as regulated.
Read more →Tell us what the problem looks like.
Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.