Every deck about AI in commercial analytics is built on the same four-rung ladder: descriptive, diagnostic, predictive, prescriptive. It has been on the first slide of that deck for fifteen years. It is not wrong, exactly. It is just too coarse to make a single decision with, and its coarseness is expensive in a specific way: it puts a correlation and a randomised holdout in the same box and calls both of them prescriptive.
This is part one of three. It is the boring part - vocabulary and a frame - and everything practical in parts two and three depends on it holding.
What the work actually is
The useful reframe is not “which model do we build”. FMCG commercial analytics operates over one very high-dimensional cube:
SKU × store/customer × region × week × channel × promo × price × distribution × availability × shopper segment
No human can enumerate that. There are more slices than there are working hours in a year, and the interesting ones are not where you would look. What machine learning is genuinely good at here is not proving causality and it is not chatting with your data. It is the systematic management of commercial attention: finding where performance changed, what it moved with, where an opportunity is sitting unclaimed, roughly how big it is, and which actions are worth testing first.
The largest near-term value of AI/ML in this domain is not “the machine proved causality.” It is that the machine decides what a scarce analyst looks at on Monday morning.
That single move converts the conversation from which model to which decision questions, which is the only version of the conversation a commercial director can participate in.
The Evidence & Action Ladder
Here is the replacement for the four-rung ladder. Ten rungs, each with its own question, its own output, and - the part that matters - its own forbidden overclaim.
| # | Rung | The question | FMCG example | Allowed phrasing | Forbidden overclaim |
|---|---|---|---|---|---|
| 1 | Observe | What happened? | Sales −8%, share −1.2pp, availability 82% | Observed fact | Any “why” |
| 2 | Localize | Where exactly? | 42% of the decline in Retailer A / South / SKU C | Localized issue | ”Retailer A caused the decline” |
| 3 | Associate | What moved with it? | Sales gap associated with availability gap and lost distribution | Associated driver hypothesis | ”Availability cost us X units” |
| 4 | Size | What order of magnitude? | Missed sales ≈ £300k | Estimate, conditional on baseline | ”We will recover £300k” |
| 5 | Predict | What is likely next? | Out-of-stock risk in two weeks | Forecasted risk | ”It will go out of stock” |
| 6 | Simulate | What does the model say if X changes? | Discount 15% → margin £1.3m | Model-implied scenario | ”Raising availability will yield Y” |
| 7 | Recommend | What do we do first? | Fix the top 100 availability issues | Recommended action under a rule | ”This action returns exactly £X” |
| 8 | Validate | Did the action work? | Post-intervention recovery vs baseline | Validated action | - |
| 9 | Causal estimate | What was the incremental effect? | Difference-in-differences promo uplift +6% | Estimated under stated assumptions | ”Proven, universally” |
| 10 | Optimize | How do we allocate a scarce resource? | Field team visits 80 stores - which? | Optimized portfolio | ”The optimizer created the effect” |
Two things about this table are worth more than the table itself.
Localization and association are not warm-up exercises. In most published maturity models, rungs 2 and 3 are the tedious prelude before the real work begins. In FMCG they are where the majority of realisable value lives, because the cube is large and nobody is currently searching it systematically. A team with no data science function should be aiming to run rungs 1–7 well, permanently, rather than aiming to reach rung 9 badly.
The forbidden-overclaim column is the actual product. A finding without an evidence label is not a finding, it is a rumour with a chart attached. The label is what stops a rung-3 association being presented to a board as a rung-9 effect, which is the single most common way analytics loses its credibility - not by being wrong, but by being right at a lower rung than it claimed.
One open question, two directions of travel
Above all ten rungs sits one question that is never answered by a dashboard:
What is the state of the business?
It decomposes into tracks, and the tracks split on a distinction that gets collapsed constantly:
| Track | How you enter | Main question | Output |
|---|---|---|---|
| State of business | Open scan | What is going on at all? | Executive radar: risks and openings |
| Issue investigation | A known event, top-down | Why did sales / share / margin fall? | Issue pack + next action |
| Opportunity finder | Open search, bottom-up | Where is the upside when nothing is broken? | Opportunity backlog, whitespace list |
| Missed sales finder | Demand existed, was not served | Where did we underserve on availability or listing? | Missed-sales action queue |
| Promo / RGM decision support | A decision is due | Which promo, price or pack? | Scenario + recommendation |
| Causal validation | Escalation for expensive decisions | Can we say X caused Y? | Causal estimate memo |
Issue investigation is top-down: the event is known, and you are narrowing. Opportunity finding is bottom-up: there is no event, and you are scanning a universe. These are different modes of thinking with different failure modes, and merging them into one process - which most “AI insights” tooling does - produces something that is mediocre at both. Part two takes each apart.
The words people keep swapping
Half the arguments in this field are vocabulary arguments that nobody has noticed are vocabulary arguments.
| Term | What it means | What it does not mean |
|---|---|---|
| Localization | Finding where a change is concentrated | It does not identify a driver |
| Slice finding | Automated search across dimension combinations for unusual behaviour | A slice is a subgroup, not a cause |
| Association | A variable is statistically linked to the target, or helps explain or predict it | Not a causal effect |
| Driver diagnosis | Business interpretation of associated signals into a likely explanation | Still a hypothesis |
| Prediction | Estimating a future value or risk under current patterns | It does not choose an action |
| Scenario / what-if | Recomputing a model with a changed input | The coefficient still has to come from somewhere |
| Prescription | A recommended action | Can be strong or very weak on evidence |
| Causal effect | The estimated effect of an intervention, do(X) | Requires a design, not just a model |
| Optimization | Choosing the best set of actions under constraints | It does not create the coefficients it optimises over |
If a team adopts nothing else from this piece, adopt this table and put it on a wall. Most of the expensive mistakes downstream are one of these nine confusions wearing a suit.
”Prescriptive” is a spectrum, not a rung
This is the correction I would most like to see propagate. In vendor language, prescriptive analytics absorbs forecasting, what-if, regression, scenarios, recommendations, optimization and causal prescription into one word. In practice it is a spectrum with an order of magnitude of rigour between its ends:
| Kind of prescription | Rests on | Example | Rigour |
|---|---|---|---|
| Rule-based | A business rule | Availability <85% on a top SKU → fix it | Low, but genuinely practical |
| Threshold / exception | An SLA or limit | Distribution −5pp → check the listing | Low |
| Association-based | Correlation, regression, SHAP | Sales gap associated with availability gap → check availability | Medium; good enough for triage |
| Forecast-based | Predicted risk | Out-of-stock risk → replenish | Medium |
| Scenario / what-if | A coefficient plus a formula | Discount 15% → best margin | Entirely dependent on the coefficient |
| Simulation-based | Response curves plus assumptions | Simulating a promo calendar | Medium to high |
| Score-based prioritisation | Value × confidence × feasibility | Top 100 store-SKU actions | Very practical |
| Optimization-based | Expected value under constraints | Allocating field-force visits | High, for decision support |
| Causal-effect-based | DiD, BSTS, DML, holdout | Availability fix → +6% | High |
| Experiment-backed | Randomised controlled test | Rollout after a holdout | Strongest available |
| Adaptive | A feedback loop | The system learns from each action | Mature |
The frame that makes this usable:
Prescription = action logic + expected value + constraints + evidence label
Drop the fourth term and the whole thing degrades into advice. The failure starts exactly where a recommendation derived from an association gets sold as a proven incremental effect - and it is usually sold that way not out of dishonesty but because the vocabulary offered no way to say the difference out loud.
Where the coefficient comes from
This is the layer I missed on my first pass through the problem, and it turned out to be the load-bearing one. Any what-if or any prescription with a number attached requires a response coefficient: how much does Y move when X moves. And the crucial realisation is that the method does not determine the strength of the claim - the design of the question and the origin of the coefficient do.
| Coefficient source | Example | What to call the output | Risk |
|---|---|---|---|
| Expert assumption | ”+10% discount → +20% units” | Assumption-based scenario | Subjective |
| Historical average | Mean uplift of similar past promos | Historical scenario | Confounded by season, support, stock |
| Correlation | Availability and sales move together | Association-based hypothesis | Confounding, reverse causality |
| Regression coefficient | Sales ~ availability + price + promo + controls | Model-based association | Not causal without a design; multicollinearity |
| Predictive ML / SHAP | Model predicts, SHAP highlights features | Predictive scenario | Feature importance ≠ causal effect |
| Quasi-experiment | Treated vs comparable control, before and after | Causal estimate under assumptions | Needs a real design |
| Randomised holdout | Random test and control stores | Experimental effect | Expensive, slow, external validity |
| Vendor benchmark | External elasticity | Benchmark scenario | May not transfer |
| Feedback loop | Real outcomes of past actions | Adaptive estimate | Requires clean action logging |
Read down that table and the point lands: a what-if is not evidence. Its quality is entirely inherited from the row it drew its coefficient from. The same regression can be an association, a prediction, a scenario or a causal estimate depending on how the question was designed - and nothing in the output distinguishes the four. Only the design does, and only a human records it.
What part one buys you
Nothing yet. That is the honest answer. This part builds no model, finds no opportunity and saves no money.
What it does is make the next two parts possible to write without lying. Part two builds the engine - the three finders, the grain they need, and the exact point where a BI tool stops being able to help. Part three asks who has to be in the room for any of it, what it costs, and why the first project should not be a vendor implementation.