A practitioner's framework · CPG × Retail × AI
Where AI actually works in the CPG & retail value chain.
Everything downstream of the factory gate is a commercial decision: how much to move, where to place it, what to charge, how to sell it, how to keep the customer. AI vendors claim all of it. Some of those claims are real.
This framework walks the chain stage by stage: the problem that has always defined it, the playbook of interventions available, and the honest read — an operator's judgment on where AI earns its keep and where it's theatre. Each stage closes with the test: one line that separates the two.
One test runs through every stage: the AI narrates; deterministic systems compute. Zero invented numbers.
The chain · factory gate onward
This framework deliberately starts at the factory gate. Research, sourcing and manufacturing are upstream of where I operate — I'd rather be useful about the half of the chain I know than complete about the half I don't.
↺ the customer signal at the end feeds back to how much you plan, buy and move · think customer backwards
Stage 01
Plan & Distribute
Demand planning, inventory, warehousing, logistics — getting finished goods to where they sell.
The problem
Every business in this chain runs on a guess. The forecast is wrong — the only questions are by how much, and in which direction. Too high, and working capital sits in warehouses ageing toward markdown or expiry. Too low, and the shelf goes empty precisely when demand showed up — and in most categories a lost sale is simply lost, not deferred. The damage compounds downstream: a small error at the plan level becomes a bigger one at the depot and a stockout at the store, because every echelon adds its own buffer to a number it doesn't trust.
Meanwhile the signal that would fix this — what actually sold, today, at the shelf — exists, but it lives in someone else's system, arrives late, filtered, and aggregated past the point of usefulness. So planning runs on shipments (what we pushed) rather than consumption (what people bought), and the whole chain plans against an echo of demand rather than demand itself.
The playbook
- Plan from consumption, not shipments. Sell-out and point-of-sale signal, however imperfect, beats sell-in data for seeing true demand. Everything else improves when the input does.
- Segment before you forecast. No single method fits a whole catalogue. Separating fast, predictable movers from slow and erratic ones — and planning each class differently — buys more accuracy than any clever algorithm applied uniformly.
- Set inventory across echelons, not node by node. Buffers sized locally always over-insure in aggregate. Multi-echelon optimisation places stock where it protects service at the least total cost.
- Sense demand short-term; plan it long-term. Blending near-term signals — weather, promotions, local events, channel shifts — onto a baseline forecast catches the swings a monthly cycle misses.
- Design the network for the demand you have. Depot footprints outlive the demand patterns that justified them. Periodic network redesign recovers cost that daily operations never will.
The honest read
This is the stage where AI has the strongest claim in the entire chain — and the one where the claim most needs auditing.
The honest case: forecasting is a pattern-recognition problem, and pattern recognition is what machine learning actually does. Models that learn seasonality, promotion response and channel behaviour across thousands of SKUs genuinely outperform the classical methods — not by magic, but because they ingest more signal than a planner can. Where the data discipline exists, this is real, compounding value. AI can also read the unstructured edges — a distributor's email, a weather alert, a competitor's launch — and flag what a numbers-only system would miss.
The hype to resist: "the AI predicted it" is not an explanation, and an unexplainable forecast is an unadoptable one — planners will quietly revert to their spreadsheets the first time the black box embarrasses them. Autonomous replenishment that no human can interrogate fails not on mathematics but on trust. And no model rescues bad master data; most "AI forecasting" disappointments are data-hygiene failures wearing an algorithm's clothes.
The test
Let the model find the pattern and explain it — why this SKU, why this week, what changed. Let the arithmetic of safety stock, replenishment and allocation run in deterministic code, auditable line by line. A forecast you can interrogate gets adopted; a number that appeared from nowhere gets ignored — or worse, obeyed.
Stage 02 · The seam
Trade & Channel
Where brand sell-in meets retail sell-through: trade terms, promotions, joint plans, and the data both sides argue over.
The problem
This is the most contested ground in the chain, and the most expensive. A fifth to a quarter of a CPG company's revenue goes into trade — promotions, terms, schemes, margins — and most companies cannot say with confidence which of those dollars grew the category and which merely moved a purchase two weeks earlier, funded a pantry load, or cannibalised their own adjacent SKU.
The brand plans on sell-in; the retailer lives on sell-through; the distributor in between absorbs the difference and tells each side what it wants to hear. Both parties hold half the data the other needs, and share it late, reluctantly, or not at all. So promotions get judged on gut and anecdote, joint business plans get negotiated on last year plus ambition, and the same arguments repeat every quarter — because nobody can prove anything.
The playbook
- Judge promotions on true lift. Baseline modelling that separates genuine incremental volume from timing shifts, pantry loading, and cannibalisation — so ROI means something.
- Share the data both ways. POS and inventory flowing brand-ward; plans and launch intent flowing retail-ward. Collaboration frameworks — joint planning, shared forecasts, vendor-managed inventory — exist precisely to institutionalise this.
- Run the distributor as a business. Distributor economics — their ROI, working capital, fill rate — treated as a strategic KPI, not an afterthought. A distributor who makes money carries your line properly.
- Simulate before you spend. What-if modelling of promo calendars — depth, timing, mechanic, channel — before committing the budget, not post-mortems after it's gone.
The honest read
Trade is where the largest sums meet the weakest measurement — which makes it simultaneously AI's biggest opportunity in this chain and its biggest bluff.
The honest case: isolating true promotional lift is a causal-inference problem with confounders everywhere — seasonality, competitor activity, placement, weather. This is exactly the class of problem modern modelling handles better than the spreadsheet-and-instinct method it replaces. Done properly, it changes the conversation between brand and retailer from opinion to evidence: this mechanic, in this channel, at this depth, returns; that one never has. AI can also make collaboration cheaper — reading the retailer's portal data, reconciling claims, drafting the joint plan — the connective work that used to take an analyst's week.
The hype to resist: a dashboard is not a decision. Much of what sells as "trade promotion AI" is descriptive reporting with a chat window — it tells you what happened, prettily, and leaves the hard question (what should we do differently?) untouched. Be equally suspicious of lift numbers produced by models nobody can explain to a customer: a promotion ROI you can't defend across the negotiating table is worthless, because the negotiating table is where it has to survive.
The test
The model estimates lift and shows its work — baseline, confounders, confidence. The trade budget arithmetic stays deterministic, because numbers that get argued over between companies must be reproducible. The seam runs on trust; an unexplainable number spends trust rather than building it.
Stage 03
Assort & Buy
What to carry, how deep, and what to drop — range architecture, open-to-buy, SKU rationalisation.
The problem
Ranges only ever grow. Every launch, every brand extension, every "just list it and see" adds a SKU; almost nothing removes one, because every SKU has a defender — the brand manager who launched it, the account that once ordered it, the customer someone remembers asking for it. The result is a long tail that consumes warehouse slots, working capital, planning attention and shelf space while a fraction of the catalogue does the actual work.
The classic cure is a rationalisation exercise: a heroic one-off cull, announced with conviction, that quietly regrows within eighteen months — because the system that produced the bloat was never touched. Buying has the mirror problem: open-to-buy committed on last season's pattern, depth spread evenly to be safe, and markdowns absorbing the difference between the buy and the truth.
The playbook
- Charge for complexity. Every SKU carries a real cost — slots, capital, attention. Make that cost explicit and the "keep everything" default collapses on its own.
- Rationalise continuously, not heroically. A standing review with clear rules — contribution thresholds, substitutability, role in the range — beats a purge every three years.
- Assort locally. The right range differs by store cluster, channel and market. One national planogram is a compromise nobody's customer chose.
- Buy to sell-through, not to sell-in. Depth allocated to demonstrated velocity; markdown discipline planned at the moment of purchase, not discovered at season's end.
The honest read
The honest case: assortment lives or dies on substitutability — if we delist this, where does the demand go? — and that question is genuinely hard for humans and genuinely tractable for models that learn purchase-affinity patterns across baskets. AI is also strong at the classification grunt-work underneath: enriching thousands of SKUs with consistent attributes so the range can be analysed at all. Attribute enrichment is unglamorous and transformative — most assortment analytics fail on dirty product data before any algorithm gets a chance.
The hype to resist: "AI-optimised assortment" pitched as a black box that hands you the perfect range. Range decisions encode strategy — which customer you serve, which fights you pick, what the brand stands for. A model can tell you one SKU's demand transfers cleanly to another; it cannot tell you whether carrying the slow import is the point of your store. Delegating the judgment along with the arithmetic is how a range becomes coherent to a spreadsheet and incoherent to a shopper.
The test
The model quantifies — transfer rates, contribution, tail cost — and narrates the trade-off. The keep/kill call stays with a human who owns the strategy; the open-to-buy arithmetic stays deterministic. AI informs the range — it doesn't get to be the range.
Stage 04
Merchandise & Price
Price and pack architecture, promotion mechanics, and the content that makes product sell across channels.
The problem
Price is the most powerful lever in the P&L and the most casually pulled. Most businesses price by cost-plus habit, competitor mimicry, or the founder's instinct — then wonder why margin leaks. Pack-price architecture drifts until the range accidentally teaches customers to buy the discounted size. Promotions train the market to wait.
And in digital channels a second problem has grown beside the first: the content burden. Every SKU now needs titles, descriptions, attributes, images and search terms — per channel, per market, per language — and merchandising teams drown in production work that has nothing to do with judgment. The result is a strange inversion: the highest-leverage commercial decisions (price) made with the least analysis, and the lowest-leverage work (content formatting) consuming the most hours.
The playbook
- Price from elasticity, not habit. Knowing which SKUs can carry price and which can't — by channel, by market — is the fastest margin lever in this chain.
- Architect the price-pack ladder. Sizes, multipacks and tiers designed so every step trades the customer up or protects the floor — not an accident of launch history.
- Promote by design, not calendar. Mechanics chosen for what they do — trial, volume, clearance — with the discipline to measure each against intent.
- Industrialise content. Product content produced as a pipeline — generated at scale, checked against a source of truth, versioned per channel — instead of artisanal copywriting under deadline.
The honest read
The honest case, in two halves. On content, generative AI is simply the right tool: producing consistent, search-ready product copy across thousands of SKUs is exactly what it does well, and the economics are not close — days instead of quarters. The discipline that matters is the gate: generation constrained by verified product data, human review where claims carry risk, and no invented specifications, ever. A hallucinated ingredient list isn't a quality issue; it's a liability. On pricing, models genuinely help estimate elasticity and simulate scenarios — where the data is dense enough, which usually means high-velocity SKUs in digital channels.
The hype to resist: fully autonomous "dynamic pricing" in contexts where customers see and remember prices. What works for airline seats corrodes trust in a grocery basket. The gap between "the model can reprice hourly" and "your customer forgives being repriced hourly" is where brands get damaged. Price is a promise as much as a number; algorithms don't carry promises.
The test
Generated content passes a deterministic fact-gate — every claim traceable to the product master. Price models recommend and explain; a human owns the price the customer sees. Publish either unreviewed, and you've handed the brand's voice and promise to a system that has neither.
Stage 05
Sell
The point of sale itself — general trade, modern trade, DTC, marketplaces — and the omni layer that unifies them.
The problem
Nobody runs one channel anymore, and the channels don't cooperate by default — they fight. The brand.com store resents the marketplace's discounts; the marketplace cannibalises the retailer; the distributor watches DTC with suspicion; the store blames online for its traffic. Each channel carries its own economics, its own operating rhythm, its own version of the customer — and left alone, each optimises itself at the portfolio's expense.
Meanwhile the operating load multiplies: every channel added is another set of content requirements, campaign calendars, fee structures, algorithms and reconciliations. Most businesses respond by staffing each channel as a silo, which locks the fighting in. The portfolio question — which channel does which job, at what contribution, for which customer — goes unasked, because everyone is too busy running their own lane.
The playbook
- Assign each channel a job. Scale, growth, profitability, discovery — channels earn their place by role in the portfolio, not by revenue alone. Contribution, not GMV, is the scoreboard.
- Unify the inventory. Stock visible and sellable across channels — store inventory serving online demand, online catalogue extending the store — is the single structural advantage of running omni properly.
- Operate the algorithms. Marketplaces are algorithmic environments; search rank, retail media and campaign mechanics are operating disciplines, not marketing afterthoughts.
- Make the portfolio legible. One view of performance across channels, on comparable economics, so the trade-offs are visible before the quarter ends rather than after.
The honest read
The honest case: selling is where AI's operational value is most immediate, because the work is high-frequency and rule-rich. Campaign bids, search terms, retail-media allocation, listing optimisation — these are decisions made thousands of times a week under measurable feedback, which is the environment where machine assistance genuinely compounds. AI also collapses the channel-proliferation tax: one product truth generating channel-specific content, one team operating what used to need five.
The hype to resist: "AI-powered growth" pitched as a substitute for channel economics. No optimisation layer rescues a channel whose contribution math doesn't work — it just spends the marketing budget more elegantly on the way down. And be careful with autonomous bidding given real money and a vague objective: an algorithm told to maximise revenue will happily buy it at negative margin, at scale, without ever technically being wrong.
The test
The machines operate inside guardrails a human set — margin floors, budget caps, brand exclusions — and every automated decision is loggable after the fact. The AI runs the repetitions; the operator owns the economics. The moment those invert, the channel is running you.
Stage 06
Serve & Retain
Service, CRM, loyalty, voice-of-customer — the relationship after the first sale, where the margin actually lives.
The problem
Acquisition gets the budget; retention earns the money. Almost every business in this chain knows the arithmetic — the repeat customer costs a fraction to serve and contributes multiples over their life — and almost every one still behaves as if the first purchase were the finish line. The CRM sends what the calendar says, not what the customer's behaviour says. Loyalty programmes pay margin for transactions that would have happened anyway. Service is run as a cost centre to be minimised rather than the highest-signal listening post the business owns.
And the voice of the customer — thousands of surveys, reviews, tickets, chats — arrives as unstructured text that nobody has time to read, so it gets compressed into a score, and the score gets reported, and nothing changes. The signal was there all along; it was just in prose.
The playbook
- Segment by behaviour, act by segment. Value tiers and lifecycle stages with genuinely different treatment — investment where the future value is, not where the noise is.
- Automate the lifecycle. Welcome, replenishment, win-back, lapse-prevention journeys triggered by what the customer does, not when the campaign calendar fires.
- Design loyalty around margin. Reward the behaviour you want more of; stop paying for the behaviour you'd get anyway.
- Mine the verbatims. The open-text feedback — reviews, tickets, survey comments — read systematically and routed to the function that can act on it.
- Serve to retain, not just to resolve. Service quality measured by what happens to the relationship afterward, not by handle time.
The honest read
The honest case: this stage holds the most underexploited AI opportunity in the chain, and it's the unstructured text. A thousand open-ended survey answers, a quarter's support tickets, a year of reviews — AI reads all of it, finds the patterns, and turns "our NPS dipped" into "customers in this segment are churning over this specific failure." That is signal businesses have always owned and never used. Done privacy-first — the analysis coming to the data, not the data leaking to the analysis — it converts the service function from cost centre to sensing organ. Service copilots that draft responses and retrieve context are similarly real: they make good agents faster without pretending to be the relationship.
The hype to resist: the fully automated relationship. A bot that deflects a frustrated customer hasn't reduced cost; it has spent the relationship to hide the cost somewhere the dashboard doesn't look. Same for "hyper-personalisation" that crosses from attentive to unnerving — the customer who wonders how do they know that? is not feeling loyal. Retention is trust economics, and trust doesn't automate.
The test
AI reads, summarises, drafts and flags; humans own the moments where the relationship is at stake. Every claimed pattern traces to real verbatims someone can pull up and read. The AI narrates what a thousand customers said — it doesn't invent what they meant.
The closing lens
An honest AI maturity ladder
Where does your business actually stand? One ladder for the whole chain — locate yourself, then look one rung up.
Decisions on experience and spreadsheets. Data exists but arrives late, aggregated, and argued over. Most of the chain, in most markets, is here — and pretending otherwise is where bad AI programmes start.
The transactions flow without retyping: orders, claims, reconciliations, reporting. Unglamorous, essential, and the honest prerequisite — automating a broken process just breaks things faster.
The business measures what's true: real lift, real elasticity, real cost-to-serve, real churn drivers. Decisions still human, but argued from evidence. Most businesses that claim to be "doing AI" are actually here — and it's a fine place to be, briefly.
Not "the algorithm decides." AI reads what humans can't — the scale, the text, the patterns — narrates what it found and why, and drafts the response, while every number the business acts on is computed deterministically and can be audited line by line. Machines handle the repetition inside human-set guardrails; people own the judgment, the strategy and the relationships. The AI narrates; deterministic systems compute; zero invented numbers.
The uncomfortable truth about the ladder: you cannot skip rungs. AI-native ambitions on manual-rung data produce expensive demos. The fastest route to Rung 4 is unglamorous work on Rungs 2 and 3 — which is precisely why so few get there.