Category
Management &
Artificial
Intelligence

Walk into any supermarket and you are standing inside a portfolio of businesses. The coffee aisle is one business. Oral care is another. Chilled ready-meals, pet food, laundry, salty snacks — each is a distinct enterprise with its own customers, its own competitors, its own economics, and its own reason to exist inside the store. Category management is the discipline of recognising that fact and acting on it: of treating each product category as a strategic business unit with its own role, strategy, targets and accountability, rather than as an undifferentiated stretch of shelving to be filled.

That single shift — from managing brands or SKUs in isolation to managing the category as a coherent whole — is the intellectual core of the discipline. It changes the question a buyer asks. Not “how many cases of this brand can I sell?” but “what is this category for, who is it for, and how do I grow the whole of it?”

DefinitionsFour ways to say the same thing

There is no single canonical definition of category management, but the major ones converge. It is worth seeing them side by side, because each emphasises a different facet of the discipline.

Pull the four definitions apart and the same four ideas appear in each. First, the category — not the brand or the SKU — is the unit of management. Second, it is run as a strategic business unit: with a role, a P&L logic, targets, and someone accountable for them. Third, every decision is anchored in shopper and consumer value, not in internal convenience. Fourth, it is executed collaboratively, between a retailer and one or more of its suppliers. Hold those four ideas; the rest of this guide is commentary on them.

First principlesThe consumer and the shopper are not the same person

One distinction deserves to be settled immediately, because the entire discipline — and every AI prompt later in this guide — depends on it. The consumer is the person who uses or eats the product. The shopper is the person standing in the aisle, or scrolling the app, making the purchase decision. They are frequently different people — the parent buying for the child, the partner doing the weekly shop — and even when they are the same person, they are in a different mindset. Category management lives at the moment of the shopping decision. It is shopper-facing by construction. Marketing builds the brand in the consumer’s head; category management wins the few seconds in front of the shelf.

Why it existsThe problem category management solves

Before category management, retail buying was organised around suppliers and brands. A buyer negotiated with Brand A, then with Brand B, and the “category” was simply whatever those separate negotiations happened to produce. Three pathologies followed. Assortment sprawled, because every supplier pushed every line and nobody owned the total. Shelf logic was incoherent, because the layout reflected supplier deals rather than how shoppers chose. And nobody was accountable for the category’s overall health — growth, margin, shopper satisfaction — because no one was looking at the category as a whole.

Category management is the cure for all three. It installs an owner, a definition, a role, a strategy and a scorecard for each category — and it does so collaboratively, so the supplier’s deep category expertise is harnessed rather than merely resisted.

Two disciplines, one nameA warning before you go further

The phrase “category management” names two distinct professional disciplines. This guide is overwhelmingly about the first — retail / FMCG category management, the sell-side discipline born in groceries. But there is a second, equally real discipline called procurement category management: the buy-side practice of grouping an organisation’s purchasing spend into categories and sourcing each strategically. Same name, opposite end of the value chain. Chapter 7 is devoted to it. For now, simply know that the ambiguity exists, so you are never confused by it.

What good looks likeThe signatures of a well-managed category

You can recognise a well-managed category without seeing a single spreadsheet. It has a clear definition that matches how shoppers think, not how the buying desk is organised. It has been assigned an explicit role in the store. Its assortment, pricing, shelf layout and promotions are mutually coherent and visibly serve that role. It is measured against a scorecard that someone owns. And it is reviewed on a cycle, so it adapts. The chapters that follow give you each of those instruments in turn — and Part II hands the cycle to a team of AI agents.

Disciplines have origin myths; category management has an origin mechanism. It could not exist until retail became data-rich, and retail did not become data-rich until products could be scanned. The story therefore begins not with a theorist but with an engineer, and a 12-digit code.

The enablerThe barcode makes the category measurable

In 1973 the IBM engineer George Laurer developed the Universal Product Code — the 12-digit numeric barcode. On 26 June 1974, a pack of Wrigley’s Juicy Fruit gum was scanned at a Marsh Supermarket in Troy, Ohio: the first commercial UPC scan. Through the 1980s, chains rolled out scanning nationally. The consequence was profound and is easy to undervalue. For the first time, a retailer knew — at the level of the individual item, store and day — exactly what had sold. Retail had become a measured environment. You cannot manage a category as a business unit until you can see its numbers; the barcode is what let you see them.

The originatorBrian Harris and the category as a business unit

The conceptual leap belongs to Brian F. Harris, an Australian-born marketing academic at the University of Southern California. In the 1980s Harris built the Apollo Space Management System — early software that calculated optimal shelf-space allocation, and the direct ancestor of every planogramming tool in use today. From that work grew a bigger idea: that a retailer should treat each category the way a corporation treats a division — with a role, a strategy, a P&L and an owner. In 1990 Harris founded The Partnering Group, the consultancy that would codify the method. The exact birth year of “category management” is genuinely fuzzy: the concept dates to the late 1980s, while the formal term and the codified eight-step model were popularised through the mid-1990s. It is most honest to describe it as a discipline developed across roughly 1989–1997.

The proofProcter & Gamble and Walmart

Two corporate moves turned theory into practice. First, in 1987, Procter & Gamble abandoned its 56-year-old brand-management structure — the very system Neil McElroy had invented at P&G in 1931 — and reorganised into category business units, each with its own sales, finance and product-development functions. The largest CPG manufacturer in the world had just declared that the category, not the brand, was the unit that mattered. Second, beginning in 1987–88, P&G and Walmart built the era’s defining collaboration: continuous replenishment, shared data, and ultimately Walmart’s Retail Link system, which gave suppliers item-store-day POS data. Walmart–P&G became the textbook proof that retailer and supplier sharing data could grow a category for both. It is still cited that way.

The crisisECR forces the method onto the industry

In the early 1990s, US grocers felt mortally threatened by mass merchandisers, warehouse clubs and drugstores piling into food. Trade associations commissioned a study, and in 1993 Kurt Salmon Associates published Efficient Consumer Response, claiming the grocery supply chain could be re-engineered to remove a figure widely quoted as up to $30 billion in cost. ECR launched the same year as an industry-wide movement, built on four pillars: efficient assortment, efficient replenishment, efficient promotion and efficient product introduction. Category management became the demand-side strategic process within ECR. ECR Europe followed in 1994; the 8-Step Approach to Category Management was launched under the ECR banner in 1996. A method that began as one academic’s framework was now an industry standard.

The measurersNielsen, IRI and the syndicated-data industry

A discipline that runs on data needs data suppliers. Two firms came to dominate. AC Nielsen evolved from store audits into scanner-based measurement; Information Resources, Inc. (IRI), founded in Chicago in 1979, pioneered scanner panels and the InfoScan national tracking service, forcing the whole industry onto syndicated scanner data. That data — standardised, cross-retailer, comparable — became the analytical fuel of category assessments and scorecards. The lineage still matters: Nielsen’s retail arm is today NielsenIQ; IRI merged with The NPD Group in 2022 to form Circana. When Chapter 10 discusses the “context” you must feed an AI, syndicated panel data from these firms is one of its load-bearing inputs.

The critiqueWhy the classic model started to crack

By the 2000s the eight-step process was under fire from its own practitioners. It was slow and resource-heavy — one widely cited industry survey found only around 9% of supplier companies used the full eight-step process; most used trimmed-down versions. It had drifted toward being retailer-centric, a lever for extracting slotting fees and captain payments rather than a genuine win-win. And it was criticised as insufficiently connected to the actual shopper. An FMI / Deloitte study reportedly found essentially every retail and CPG respondent believed the prescribed process needed to change. The response was two-fold: streamlined proprietary processes (Nielsen, for instance, promotes a five-step version), and a decisive turn toward shopper-centric category management — the subject of Chapter 4.

If Chapter 1 gave you the philosophy and Chapter 2 the history, this chapter gives you the machinery. Almost everything an AI agent does in Part II of this guide is an automation, an acceleration or an augmentation of one of the four instruments described here: the eight-step process, the category-role framework, the category-captain model, and the four tactical levers. Learn them as a practitioner would; you cannot supervise an agent doing work you do not understand.

The processThe eight-step category management cycle

Developed by The Partnering Group and codified through ECR, the eight-step process is the canonical reference model. It is a cycle, not a project: step 8 loops back to step 1. Few organisations run all eight steps in full — most use streamlined versions — but every streamlined version is a compression of these eight, so this is the model to know.

The rolesFour jobs a category can hold

Step 2 deserves its own discussion, because it is where scarce resources — space, capital, price investment, management attention — are deliberately allocated rather than spread evenly. A retailer assigns each category exactly one of four roles, based on how important the category is to the target shopper, how much it differentiates the retailer, the competitive dynamics, and its financial contribution.

The discipline of roles is that everything else follows from them. A Destination category warrants deep assortment, aggressive known-value-item pricing, premium space and a strong promotional calendar. A Convenience category warrants the opposite on every lever. A category plan whose tactics contradict its assigned role is, by definition, a broken plan — and, as Chapter 13 notes, “does this plan match its role?” is exactly the kind of check an AI governance agent can run automatically.

The collaborationThe category captain — and its dangers

Retailers rarely have the analytical firepower to manage every category deeply in-house, so they borrow it. A category captain is a single supplier — almost always the category’s leading manufacturer — that a retailer designates to provide deep analysis and recommendations for the whole category, including the products of the captain’s own competitors. In exchange for privileged data access, the captain invests its shopper insight and category expertise. A second supplier is often named category validator to challenge the captain’s recommendations; a neutral, brand-agnostic category advisor may also be used.

The arrangement is genuinely useful and genuinely dangerous. The danger is structural: you have asked a self-interested competitor to advise on how its rivals are treated on the shelf. Antitrust scholars and the US FTC have identified two principal harms — competitive exclusion (the captain quietly disadvantages rivals’ facings or recommends their delisting) and facilitated collusion (one captain serving most retailers in a market becomes a common information conduit). The accepted safeguards: captains should advise, not decide; retailers retain final authority; validators review; and the captain is never told how rivals price or promote. Hold this thought — Chapter 13 shows that an AI category captain sharpens every one of these concerns.

The leversThe four tactical levers

Step 6 executes strategy through four coordinated levers — the category-management adaptation of the marketing mix. Their golden rule is internal coherence: the four must reinforce each other and the category role.

The scorecardHow a category is measured

Steps 3 and 4 run on numbers. A category scorecard mixes growth, profit, productivity and shopper metrics. The most important to know:

• GMROI — Gross Margin Return on Inventory Investment. Gross margin divided by average inventory cost; the single best measure of how hard the category’s capital works. A GMROI of 4 means $4 of margin per $1 of inventory.
• Sales per linear metre / per facing / per square foot — space productivity; the basis of every planogram decision.
• Penetration — the share of baskets (or households) that contain the category. The clearest measure of how broadly a category recruits shoppers.
• Basket / attachment analysis — what is bought alongside what; items per basket; cross-sell rates.
• Market share & category growth — the category’s and brand’s share of a defined market, and its movement over time.
• Inventory turns, sell-through, on-shelf availability, price index, promo ROI, markdown rate — the operating vital signs.

The classic eight-step process has a quiet bias. It begins with the category — its definition, its data, its performance — and only later asks about the shopper. Shopper-centric category management inverts that. It does not throw away the eight steps; it re-anchors them, so that the shopper’s needs, missions and decision logic drive category definition, assortment, space and promotion from the first step rather than the fourth. This chapter explains the three instruments that make the re-anchoring possible — and they matter doubly here, because they are exactly the instruments an AI is best placed to build for you.

The instrumentThe Consumer Decision Tree

The Consumer Decision Tree (CDT) — increasingly called the Shopper Decision Tree, because it is the shopper, not the end consumer, who chooses at the shelf — is a graphical map of the hierarchy of attributes a shopper applies when choosing within a category. It answers a precise question: in what order do shoppers narrow their choice?

Consider hot beverages. One shopper’s tree might branch first on form (ground / beans / pods / instant), then on caffeine (regular / decaf), then on roast, then on brand, then on pack size. The order of those branches is the whole point. It tells you the shelf should be blocked by form first — because a shopper who has decided “pods” will not even look at the ground-coffee bay. It tells you which SKUs are genuine substitutes (same leaf of the tree) and which are not. And it tells you, when you rationalise the range, which items you can cut without leaving a need-state uncovered.

A terminology note worth carrying forward: vendors use “Consumer Decision Tree,” “Shopper Decision Tree” and “Category Decision Tree” more or less interchangeably. They mean the same artefact. When Chapter 11 gives you a prompt to generate one with an AI, it does not matter which name you use — it matters that you specify the attributes and the data the tree should be built from.

The instrumentShopper missions

The same shopper behaves differently on different trips. A shopping mission is defined by the motivation behind the trip, its context, and the behaviours that follow. The common missions:

• Big stock-up — the planned weekly or fortnightly shop; large basket, price-conscious, list-driven.
• Top-up / fill-in — a quick mid-week replenishment of a few items.
• Immediate consumption / food-to-go — bought to be used now; convenience and speed dominate.
• Special occasion — a dinner, a holiday, an event; higher willingness to pay, more exploration.
• Emergency / distress — the forgotten essential; the shopper is captive and price-insensitive.

Missions explain why the same category plays a different role on different trips, and why store format, assortment depth and layout must flex. They are also the bridge to Chapter 12’s synthetic personas: a well-specified mission is precisely the brief from which an AI can simulate a shopper.

The instrumentThe path to purchase

The path to purchase is the full, multi-channel journey a shopper takes from a need or trigger, through awareness, consideration and decision, to the purchase and beyond. In a single-channel world it was a short walk down an aisle. Today it threads through search, social, reviews, apps, click-and-collect and the physical shelf — and category management must influence the category’s assortment, content and visibility at every touchpoint, not just the last one. Mapping the path reveals where the category is won and lost; it is the reason the digital shelf of Chapter 6 is now as strategically important as the physical one.

Classic vs. shopper-centricWhat actually changes

Who drove itLoyalty data and the insight firms

Shopper-centric category management did not arrive as a theory; it arrived as a capability, the moment retailers could see households rather than just transactions. Tesco’s Clubcard (1995) and the firm dunnhumby turned loyalty data into household-level purchase histories — who buys, how often, what they switch to, what they never buy. Kroger’s 84.51° does the same in the US. That data is what lets a decision tree be estimated rather than guessed, a mission be sized rather than asserted. It is also, not coincidentally, the richest single input you will feed an AI in Chapter 10.

Ask ten retail professionals to draw the line between category management, trade marketing, shopper marketing and revenue growth management, and you will get ten different maps. The terms overlap, the org charts vary by company, and the vocabulary drifts. But the disciplines are distinct, and a category manager who cannot place them cannot collaborate cleanly with the teams that own them. This chapter is the map.

Push vs. pullTrade marketing and shopper marketing

These two are the easiest pair to separate, because they pull in opposite directions. Trade marketing is B2B marketing aimed at the intermediaries — retailers, distributors, wholesalers. It is a push strategy: pushing product through the channel via sales programmes and trade promotions. Shopper marketing is aimed at the shopper at or near the point of sale. It is a pull strategy: drawing shoppers to the product through in-store activation, displays and digital touchpoints.

Category management sits between them. It supplies the data-backed category rationale that makes a trade-marketing pitch credible to a retail buyer, and it shares its insight base — decision trees, missions, path to purchase — with shopper marketing. Roughly: category management decides what goes where on the shelf and why; shopper marketing decides how to activate shoppers around it; trade marketing decides how to sell the whole programme into the channel.

The collaboration wrappersJBP, ECR and CPFR

Joint Business Planning (JBP) is the commercial and governance wrapper around category plans — a multi-year, cross-functional agreement between a retailer and a supplier that turns category strategy into committed shared targets, investments and promotional calendars. If category management produces the plan, JBP is the handshake that funds and commits it.

Efficient Consumer Response (ECR), as Chapter 2 explained, is the parent movement — category management was its demand-side process. Its supply-side counterpart is CPFR (Collaborative Planning, Forecasting and Replenishment), pioneered by Walmart and Warner-Lambert in 1995 and formalised by the VICS association in 1998. CPFR has retailer and manufacturer share forecasts and jointly resolve exceptions. The clean way to hold it: category management and shopper marketing create demand; CPFR and replenishment fulfil it; JBP commits the money behind both.

The monetisation layerRevenue Growth Management

Revenue Growth Management (RGM) — also called Net Revenue Management — is a data-driven discipline for optimising net revenue and margin, not just volume, across five levers: pricing, promotions, Price Pack Architecture (PPA), trade terms, and channel mix. Price Pack Architecture is the structuring of the portfolio across price points and pack sizes to match consumer need-states. RGM effectively absorbs and systematises the pricing and promotion levers of category management and connects them to portfolio and pack decisions. Category management defines the category framework; RGM optimises how money is made within it.

The execution layerMerchandising, assortment & space planning

Retail merchandising is the umbrella discipline of what is sold and how it is assorted, priced, displayed and replenished. Within it, assortment planning (what to carry) and space planning / planogramming (how to display it) are the executional core of category management. Category management provides the strategic intent of steps 1–5; merchandising executes the tactics of steps 6–7. In apparel, as Chapter 6 shows, this same execution layer is called buying & merchandising and is governed by the Open-To-Buy budget.

The new neighbourRetail media networks

The newest discipline on the map — and the one most actively reshaping category management — is the retail media network (RMN): a retailer selling advertising inventory to brands, monetising its own first-party shopper data. US retail-media ad spend has been projected in the region of $60–70 billion a year. RMNs matter to category management for two reasons. First, they draw on exactly the category-level purchase intelligence that category teams produce. Second, and more disruptively, they invert the historical relationship: the retailer, traditionally the discerning buyer to whom the supplier sold a category-management service, now also sells media back to that same supplier. Sponsored placements increasingly shape the digital shelf the way slotting fees once shaped the physical one — reopening, in a new form, every fairness and antitrust question of the category-captain model.

The internal competitorPrivate label

Private label — the retailer’s own brands — is not a separate discipline so much as a category-management problem turned inward. Modern private label uses a tiered, “laddered” architecture: an economy tier for price-driven shoppers, a standard mid-tier for the mainstream, and a premium tier (often branded away from the retailer’s name). Each tier is managed with category-management foundations, as a sub-brand with a target segment. The hard part is that private label competes with the very national brands a category captain represents — so growing it well means modelling cannibalisation carefully and accepting that, in the hard-discount model of Chapter 6, private label simply is the category.

Category management is often taught as if it were a single, universal method. It is not. The eight-step process and the four roles are a common grammar, but each industry speaks a different dialect — because each faces a different shopper, a different shelf, a different clock. A planogram is meaningless to a fashion buyer; an Open-To-Buy budget is meaningless to a grocer; the “shelf” itself dissolves into a search ranking the moment you move online. This chapter tours ten verticals. Read it as a reminder that, in Part II, any AI you deploy must be told which dialect it is working in.

Vertical 01Grocery & supermarkets — the birthplace

Grocery invented category management and remains its most mature practitioner. A supermarket is organised into several hundred categories, each run as a P&L, each with a retailer category manager collaborating with a supplier captain. The signatures are high SKU counts, perishability, frequent promotion and deep loyalty-data integration. Tesco pioneered loyalty-driven category management through Clubcard and dunnhumby; Kroger built the same capability in the US through 84.51°; Walmart’s Retail Link gave suppliers the POS data that makes captaincy work.

Hard discount, though, breaks the model entirely. Aldi and Lidl carry well under 2,000 SKUs against a supermarket’s 30,000–50,000, and private label can be 70–90% of the assortment. There is no retailer-versus-manufacturer collaboration to speak of, because the discounter is the brand. Category management becomes ruthless in-house range curation, sourcing and product development — one category is often one or two SKUs, and the captain role is internalised.

Vertical 02FMCG / CPG manufacturers — the other side of the table

For a manufacturer, category management is the primary lever for influencing a shelf it does not own. Large CPG firms run dedicated category leadership teams that build multi-year “category visions” — growth roadmaps they bring to retailers as captain or validator. The captain provides the retailer with marketplace expertise, shopper insight, planogram recommendations and space-to-sales analysis. Procter & Gamble is a default captain candidate across many of its categories; Unilever, Nestlé, Mars, Coca-Cola and PepsiCo all run category-leadership functions. The 2025–26 wave of CPG portfolio reshaping — Mars acquiring Kellanova, Unilever spinning off ice cream, Nestlé exiting water — is itself category-portfolio management played at the corporate level.

Vertical 03Pharmacy & drugstore — two stores in one

A drugstore is really two stores. The dispensary handles regulated prescription medicine and is managed more like a procurement formulary. The front of store — OTC, health & wellness, beauty, personal care — is classic retail category management, leveraging pharmaceutical credibility to sell adjacent categories. Front-of-store profitability has been a chronic struggle: Walgreens has refocused on categories of leadership such as women’s health and a large owned-brand programme; CVS has tested small-format stores. In Germany, dm and Rossmann show how far the own-brand model can run — dm’s Balea and alverde, Rossmann’s Isana and enerBiO — with third-party quality validation used as a category-trust strategy.

Vertical 04DIY & home improvement — the line review

Home improvement organises category management around the line review: a periodic, formal re-evaluation of every vendor and the merchandising mix within a department, with vendors competing for placement on POS data, margin and supplier performance. The defining complication is two shoppers in one aisle — the DIY consumer and the Pro/trade customer — with different missions and basket economics. Home Depot and Lowe’s in the US, and B&Q, OBI and Bauhaus in Europe, all run this line-review-driven structure; fill rate, inventory turns and sales per square foot dominate the scorecard.

Vertical 05Fashion & apparel — buying & merchandising

Apparel barely uses the words “category management.” Its parallel discipline is buying & merchandising, structured as department → class → subclass and governed financially by the Open-To-Buy — a budget for what a buyer may purchase in a period. Style and seasonality dominate over the stable “category-as-business-unit” idea; the scorecard is sell-through rate, markdown percentage and weeks of cover, not long-run category share. Zara is the instructive extreme: it does not forecast finished-goods quantities at all — it forecasts fabric and uses quick response to design-to-shelf in weeks, replacing fixed category plans with continuous dynamic assortment.

Vertical 06Consumer electronics — thin margins, vendor theatre

Electronics categories — TVs, laptops, mobile, white goods, gaming, accessories — are managed amid rapid obsolescence and price deflation. Thin hardware margins push category management toward attachment: warranties, accessories, services and financing. Vendors co-invest heavily — funded displays, in-store specialists, shop-in-shop arrangements that blur captaincy. Best Buy pioneered buy-online-pick-up-in-store as a way for physical retail to compete with pure-play e-commerce; Europe’s MediaMarktSaturn is explicitly repositioning “from a product retailer to a solution-oriented omnichannel platform.”

Vertical 07E-commerce & marketplaces — the shelf dissolves

Online, the “shelf” becomes the digital shelf — search results, ranking, browse pages, recommendations — and category management splits in two. In first-party (1P) retail, a vendor manager owns the wholesale relationship: terms, purchase orders, assortment. In third-party (3P) marketplace, sellers sell direct and the retailer’s category team curates, sets policy and shapes discoverability rather than buying stock. Shelf position becomes search ranking and content quality; assortment can be near-infinite; category management merges with SEO, retail-media bidding and pricing algorithms. The KPIs change accordingly — share of search, glance views, buy-box win rate, content completeness, review rating.

Vertical 08Convenience & fuel — the forecourt takes over

A convenience store’s tiny footprint forces extreme SKU discipline and a focus on high-velocity, high-margin immediate-consumption categories. The strategic shift of the decade: fuel no longer drives growth — food does. Prepared and fresh food has become the new destination category, with the channel converging on quick-service restaurants. 7-Eleven is building larger, food-focused stores, running a delivery app with a subscription tier, and launching its own retail media network — convenience retail acquiring the full modern category toolkit.

Vertical 09B2B distribution & wholesale — category management as a service

Distributors apply category management to manage vast multi-manufacturer SKU portfolios — and increasingly sell it as a service to their business customers. Sysco, the world’s largest broadline foodservice distributor, markets “category management and insights” to restaurants and has launched a third-party marketplace, bringing e-commerce category logic into B2B. AutoZone manages 70-plus distinct categories with dedicated category-manager roles. The “shopper” here is a professional buyer, and assortment breadth is the value proposition.

Vertical 10The long tail — automotive, telecom, banking, QSR, healthcare

• Automotive parts: the unique problem is vehicle-application fitment — the same category managed against thousands of make/model/year combinations.
• Telecom: handsets are often a subsidised acquisition category whose role is to drive the high-margin connectivity and services categories; portfolios now extend into mobile financial services.
• Banking: “categories” are product lines — accounts, mortgages, cards, insurance — managed as portfolios; KPIs are products-per-customer and lifetime value, not shelf space.
• Hospitality / QSR: the menu is the assortment; menu engineering classifies items by popularity and profitability into stars, plowhorses, puzzles and dogs.
• Healthcare: procurement-side category management dominated by Group Purchasing Organisations that aggregate buying power across medical-supply categories.

Cross-cuttingThree forces that bend every vertical

Scale and market maturity. Large retailers run category management in-house with dedicated teams; small retailers receive it — planograms and advice — from suppliers, and are correspondingly more influenced by the captain. Developed markets are dominated by organised modern trade; in many emerging markets, traditional trade — kiosks, corner shops, open markets — can be the majority of consumption, and category management shifts from planograms to distributor- and rep-led execution.

The omnichannel problem. A category must now be managed across store, web and app at once — and a winning physical planogram is not a winning search-ranking strategy. Best practice fuses POS, market research, online search and e-commerce sales into one shopper view.

Retail media. As Chapter 5 noted, RMNs are turning paid placement into a determinant of the digital shelf across every vertical that has one.

Search “category management” and roughly half the results will be about retail and roughly half about procurement. They are different disciplines. Retail category management, the subject of every other chapter in this guide, is a sell-side practice: how a retailer shapes its offer to shoppers. Procurement category management is a buy-side practice: how any organisation — a manufacturer, a hospital, an airline, a bank — organises its purchasing spend. Same two words, opposite ends of the value chain. This chapter gives procurement category management its due, both to dissolve the confusion permanently and because its core frameworks are sharp enough to be worth borrowing.

The definitionWhat procurement category management is

In procurement, category management means organising the buying team’s resources to face outward onto the supply markets that serve the organisation. Spend is grouped into coherent categories — IT hardware, logistics, marketing services, raw polymer, temporary labour — and each is given a tailored sourcing strategy. The objective is not sales growth but the optimisation of total cost of ownership, supply risk and value. It is continuous, long-term and governance-led, owned by procurement rather than by merchandising.

The frameworkThe Kraljic matrix

If retail category management has the eight-step process, procurement category management has the Kraljic matrix — created by Peter Kraljic, then of McKinsey, in the 1983 Harvard Business Review article “Purchasing Must Become Supply Management.” It segments every purchase on two axes — profit impact and supply risk — into four quadrants, each with its own strategy.

The Kraljic matrix is to procurement what the category-role framework is to retail: a way of deciding, deliberately, that not every category deserves the same treatment. It also underpins supplier segmentation — managing strategic, leverage, bottleneck and routine suppliers with different intensity rather than treating all suppliers alike.

The processThe 7-step strategic sourcing cycle

Where retail has eight steps, procurement category management most often runs a seven-step strategic sourcing process:

1. Project assessment — baseline current spend, supplier relationships and stakeholder expectations.
2. Category profile — analyse spend patterns, the supplier landscape, market dynamics and risk.
3. Sourcing strategy — define the approach, using Kraljic positioning.
4. Supplier selection criteria — weighted evaluation criteria aligned to business needs.
5. Supplier engagement — run RFI / RFP / RFQ to gather proposals.
6. Negotiation — structured, data-led negotiation.
7. Implement agreements — transition contracts to operations, with monitoring and compliance.

Spend analysis — consolidating, cleaning and categorising all supplier spend into a single view — is the foundational input to the whole cycle. The Chartered Institute of Procurement & Supply (CIPS), the global professional body, publishes its own multi-phase category-management model spanning initiation, analysis, strategy and implementation. (CIPS also maintains a separate, longer procurement-and-supply cycle; the two should not be conflated.)

Side by sideThe two disciplines compared

Why it is in this guideThe borrowable ideas

Two things from procurement category management are worth carrying back. First, the Kraljic logic — segment by impact and risk, then differentiate treatment — is a clean mental model that applies to suppliers, to SKUs, and even to which categories deserve AI investment first (Chapter 14 uses exactly this kind of prioritisation). Second, procurement was an early and aggressive adopter of AI agents: “category agents” that monitor spend, flag savings opportunities and draft sourcing recommendations are already a live use case, and consultancies cite material efficiency gains from them. The agentic operating model of Chapter 9 has a cousin in the procurement world — and the two are converging.

Before a single prompt, a sober chapter. The distance between what AI can do for category management and what it reliably does today is wide, and a guide that pretended otherwise would set you up to fail. So this chapter does three things: it states the credibility problem plainly, it sorts every major capability into proven, emerging or hype, and it explains why most of the value on offer is not, in fact, generative AI at all.

The reality checkTwo numbers to keep on the desk

Two findings should anchor your expectations. On the demand side of belief: a 2025–26 McKinsey survey found that 71% of merchants said AI merchandising tools had so far had limited or no effect on their business, and 61% said their organisation was not at all, or only slightly, prepared to scale AI across merchandising — the reasons cited being fragmented systems, data too messy to generate useful recommendations, and uneven adoption. On the supply side of hype: Gartner has predicted that more than 40% of agentic-AI projects will be cancelled by the end of 2027 on cost, unclear value or weak risk controls, and has coined the term “agent washing” for the rebranding of ordinary chatbots and automation as “agents.”

Neither number says AI does not work for category management. They say something more useful: the constraint is rarely the model — it is the data foundation and the operating model around it. That is precisely why Chapters 9, 10 and 14 of this guide spend more time on architecture, context and roadmap than on models.

The honest mapProven, emerging, hype

Here is the whole landscape, sorted. The badges are deliberate: build your near-term plan on the green rows, pilot the amber, and treat the red as a direction of travel rather than a purchase.

The surpriseMost of the value is not generative AI

The single most useful thing to understand about AI in category management is that the value leaders — forecasting, optimisation, computer vision — run on classical machine learning and operations research, not large language models. Gradient-boosted trees forecast demand. Optimisation engines allocate shelf space. Deep-learning vision models read shelves. These have been quietly deployed for years.

Generative AI and LLMs sit on top of that stack as an interface and productivity layer: they draft the category review, explain a forecast in plain language, summarise a thousand product reviews, and answer questions about the data in conversation. That is genuinely valuable — it is most of what Chapter 11’s prompt library exploits — but it is augmentation, not the engine. The mental model to carry into the rest of Part II: classical ML and optimisation do the calculating; the LLM does the reasoning, drafting and conversing; the agent layer of Chapter 9 orchestrates between them.

Where genAI genuinely helps todayThe deployed copilot use cases

• Category review automation — drafting performance summaries and recommendation narratives, removing the manual report-assembly that consumes a category manager’s week.
• Conversational analytics — natural-language question-and-answer over category data, so insight does not wait on a BI ticket.
• Unstructured-data synthesis — turning reviews, social posts and survey verbatims into category themes.
• Forecast explanation — putting a plain-language “why” on a number a classical model produced.

Note the pattern: every one of these is language work wrapped around a calculation someone else did. That is the genAI sweet spot in this discipline, and the prompt library is built to exploit exactly it.

The instinct, when people imagine “AI category management,” is to picture a single, all-knowing model that ingests everything and emits a category plan. That instinct is wrong, and following it produces a system you can neither trust nor debug. The discipline already tells you the better architecture. Category management is not one task; it is eight steps, four levers and a review loop. So the AI should not be one agent — it should be a team of specialised agents, each owning a part of the process, coordinated by an orchestrator, and supervised by a human. This chapter specifies that team.

First, definitionsAssistant, agent, multi-agent system

Three words are used loosely; this guide uses them precisely.

Most organisations should start with assistants — the prompt library is usable today with nothing more — and graduate to agents and a multi-agent system as their data foundation and governance mature. Chapter 14 sequences that journey.

The architectureA nine-agent category team

The reference architecture below maps one agent to each major job in the eight-step process, plus three cross-cutting agents — shopper insight, simulation and governance — that serve all the others. An orchestrator sits above them; a human category manager sits above the orchestrator. Below everything sits the data and tool layer of Chapter 10.

The rosterWhat each agent owns — and what tools it needs

An agent is only as good as the tools it can call. The table below pairs each agent with its job and the tools it must be given access to. “Tool” here means anything the agent can invoke — a database query, a trained model, a piece of planogram software, a web search.

The control principleHuman-in-the-loop, by design

Notice what the human at the top of Figure 9.1 does, and what they do not do. They set goals, supply constraints, and approve or reject plans. They do not draft the assessment or tune the planogram. This is the human-in-the-loop (HITL) pattern: the agents do the work; the human owns the decision. A lighter variant, human-on-the-loop (HOTL), has the human monitor a running system and intervene only by exception — appropriate for low-stakes, reversible actions once trust is earned.

The rule of thumb for where to place the human: the more consequential, less reversible and more externally visible the action, the more it requires explicit human approval before execution. Generating a draft assessment is safe to automate fully. Re-pricing a Destination category’s known-value items, or delisting a competitor’s SKU, is not — those cross the Governance Agent’s desk and then the human’s. Chapter 13 makes this an explicit decision rights table.

How a cycle runsOne category review, agent by agent

1. The human sets the goal: “Run the Q3 review for the coffee category; grow margin without losing penetration.”
2. The Orchestrator loads the context pack (Chapter 10) and sequences the steps.
3. The Definition and Shopper-Insight agents confirm the decision tree and the live shopper missions.
4. The Assessment agent produces the diagnostic and ranks the gaps.
5. The Strategy agent confirms the role and proposes scorecard targets and a strategy.
6. The Assortment, Pricing, Space and Promotion agents draft coherent tactics.
7. The Simulation agent runs the proposed plan against scenarios and reports the expected range of outcomes.
8. The Governance agent checks coherence, antitrust exposure and guardrail compliance, and flags anything for escalation.
9. The Orchestrator assembles the plan; the human reviews, edits and approves.

Every numbered step above corresponds to a prompt in Chapter 11. The architecture and the prompt library are the same thing, seen twice.

Here is the most common way an AI category-management initiative fails. Someone opens a chat window, types “optimise my coffee assortment,” receives a fluent, generic, slightly wrong answer, and concludes the technology is not ready. The technology was ready. The context was missing. An LLM with no context about your category is a confident stranger who has never seen your stores — and it will answer anyway. Context engineering is the discipline of assembling, structuring and supplying everything the model needs to stop being a stranger. It is the single highest-leverage activity in Part II, and McKinsey’s finding that “data too messy to generate useful recommendations” is a top failure cause is, at root, a context-engineering failure.

The principleThree kinds of context

Useful context comes in three layers, and an effective system supplies all three.

The deliverableThe Category Context Pack

The practical output of context engineering is an artefact this guide calls the Category Context Pack — a single, structured, version-controlled document (or a small set of them) that travels with the category and is supplied to every agent and every prompt. Building it is mostly a writing and data-assembly job, and it is mostly a one-time cost per category. The template below is the recommended structure; treat the headings as mandatory and the depth as proportional to the category’s role.

# CATEGORY CONTEXT PACK
# Version: __ | Owner: __ | Last refreshed: __ | Refresh cadence: __

## 1. IDENTITY & SCOPE  [static]
- Category name and one-line definition
- Sub-segments / the shopper decision tree (attributes, in order)
- What is explicitly IN and OUT of scope
- Retailer banner, store formats, channels covered (store / web / app)
- Geography and market(s)

## 2. STRATEGIC FRAME  [static]
- Retailer positioning and target shopper(s)
- Assigned category role (Destination / Routine / Occasional / Convenience) + why
- Category vision / 3-year ambition
- The role rubric: what each role implies for each lever

## 3. SHOPPER CONTEXT  [static + dynamic]
- Shopper segments and their share of category sales
- Shopping missions in which this category is bought
- Key need-states and how well each is currently served
- Known purchase barriers and triggers

## 4. PERFORMANCE  [dynamic — refresh weekly]
- Sales, units, margin $, margin %, growth vs. prior year
- Category & brand market share and trend
- Penetration, frequency, basket size, attachment
- GMROI, inventory turns, on-shelf availability
- Scorecard: targets vs. actuals

## 5. THE FOUR LEVERS — CURRENT STATE  [dynamic]
- Assortment: SKU count, range list, productivity per SKU, private-label share
- Pricing: price ladders, KVI list, index vs. competitors, PL price gaps
- Space: current planogram, facings, space-to-sales ratio
- Promotion: recent promo calendar, depths, mechanics, measured ROI

## 6. COMPETITIVE & EXTERNAL CONTEXT  [retrieved]
- Key competitors’ assortment, pricing and known activity
- Relevant macro signals: inflation, weather, seasonality, events
- Regulatory or compliance constraints on this category

## 7. RULES OF ENGAGEMENT  [static — critical]
- Hard constraints (min/max SKUs, contractual ranging, legal listings)
- Pricing guardrails (floors, ceilings, index limits)
- Antitrust boundaries (see Chapter 13)
- Decision rights: what AI may draft vs. what needs human approval

## 8. DATA DICTIONARY & PROVENANCE  [static]
- Each data source, its owner, freshness, and known quality limits
- Definitions of every metric used above (no ambiguous terms)
- Known gaps the model must NOT guess to fill

The inputsThe data that feeds the pack

Section 4 onward of the pack is only as good as the data behind it. The load-bearing sources, and what each uniquely provides:

The disciplineFive rules of context hygiene

• State provenance and freshness. Every number should carry where it came from and when. A model that knows a figure is three months old can hedge; one that does not will assert.
• Declare the gaps. Section 8 of the pack explicitly lists what is not known. An LLM’s most dangerous habit is filling silence with plausible invention; naming the gap is the cheapest defence against hallucination.
• Define every term. “Penetration,” “share” and “margin” each have several meanings. The data dictionary removes the ambiguity so the model and the human mean the same thing.
• Right-size the pack. More context is not always better — an overstuffed pack buries the signal. Match depth to the category’s role; a Convenience category does not need a Destination category’s dossier.
• Version it. The pack is a living document under change control, not a one-off export. Stale context is worse than no context, because it looks current.

This is the working core of the guide. Each prompt below corresponds to an agent from Chapter 9 and a step from the eight-step process. They are written to be used immediately — with a capable model and nothing more than a chat window — and to scale into the multi-agent architecture unchanged, because an agent’s system prompt and a copilot’s instruction are the same artefact.

How to use this library

• Placeholders appear as {{LIKE_THIS}}. Replace every one before running — most are filled directly from the Category Context Pack of Chapter 10.
• Always attach the Context Pack. These prompts assume it is supplied. Without it you will get the confident stranger of Chapter 10.
• Run Prompt 1 first, once per session — it is the system prompt that sets behaviour for everything after.
• Keep a human in the loop. Every prompt ends by asking the model to flag assumptions and uncertainties. Read those flags; they are the point.

Set this once at the start of a session, or as the system prompt of the Orchestrator Agent. It establishes role, method, tone and guardrails for every prompt that follows.

The foundational “definition” prompt. It produces the category scope and the shopper decision tree — the artefact every later prompt depends on. Run it against POS, panel and, ideally, switching data.

The diagnostic. It reads the four classic lenses — consumer, market, retailer, supplier — and, crucially, ranks the opportunities rather than merely listing them.

Assigns — or pressure-tests — the category’s role. Best run as a challenge to a role you have proposed, so the model argues both sides.

Turns role and ambition into measurable, owned targets — the contract the rest of the plan is judged against.

Chooses the strategy archetypes that close the prioritised gaps and hit the scorecard — the bridge from diagnosis to tactics.

The first tactical lever. Produces an add / keep / cut recommendation anchored to the decision tree, so no need-state is stranded.

The second lever. Builds coherent price ladders, selects known-value items and manages private-label gaps — within explicit guardrails.

The third lever. Produces the logic of the planogram — blocking, facings, adjacencies — as a brief a space-planning tool or merchandiser then executes.

The fourth lever. Builds a promotional calendar judged on incremental, not headline, lift — and links each event to a strategy.

Designs the what-if scenarios that pressure-test a plan before it is committed — the brief the Simulation Agent of Chapter 12 then runs.

Generates a structured shopper persona for qualitative pressure-testing. Use with the caution of Chapter 12: a complement to real research, never a replacement.

The final gate. Run this against any completed plan before it reaches a human approver — it is the prompt behind the Governance Agent of Chapter 13.

Every category plan is a forecast in disguise. “Cut these twelve SKUs” assumes shoppers will switch rather than walk; “raise this known-value item 4%” assumes an elasticity; “run this multibuy” assumes an incremental lift. Traditionally those assumptions were tested the expensive way: in the actual stores, with the actual quarter, and you found out you were wrong by missing the number. Simulation moves the test before the commitment. It is the part of the AI playbook that the question “are there options for simulation runs?” points straight at — and the answer is a clear yes, across five distinct techniques of very different maturity.

Why simulateThe three jobs of a simulation

Simulation does three things a static plan cannot. It quantifies uncertainty — turning a single projected number into a range of outcomes with probabilities. It stress-tests assumptions — showing which inputs the plan’s success actually hinges on, so attention goes to the ones that matter. And it de-risks the irreversible — letting you discover that a planogram reset or a delisting fails in silico, where failure costs a compute run rather than a quarter.

The optionsFive simulation techniques

“Simulation” is not one thing. Five techniques are in use, and they sit at very different points on the proven-to-emerging scale of Chapter 8. Choose by the question you are asking.

Technique 1–2The two you should use now

What-if scenario modelling is the one to start with, because it is proven and already sits in the planning tools most retailers own. You change an input — pull a SKU, move a price, reallocate facings — and the tool re-runs the demand and financial model to show the effect on the scorecard. Run before a plan is submitted, it converts category management “from intuition to financial simulation,” in the words the vendors use. The Simulation Agent of Chapter 9 is, at its simplest, an agent that drives this capability through the scenario set that Prompt 11 designed.

Monte Carlo demand simulation adds the dimension what-if modelling lacks: probability. Instead of one demand number, you specify demand as a distribution, draw from it thousands of times, and read out the full spread of outcomes — the median, the bad tail, the probability of missing the margin target. For any plan whose success is genuinely uncertain — a new range, an aggressive price move — a Monte Carlo run is the honest way to present it to an approver: not “this delivers +3%” but “this delivers between −1% and +6%, with a 70% chance of clearing target.”

Technique 3–4The two to pilot

Virtual and VR shelf testing is different in kind: it puts real human shoppers into a photorealistic virtual store and watches them shop — navigation, hesitation, what they pick — often with eye-tracking that yields an attention heatmap of the shelf. It is the cheapest, fastest way to test whether a new planogram is actually shoppable before a single store is reset; vendors cite testing several times faster and cheaper than physical mock-ups. Treat it as a pilot-grade capability with real value for high-stakes resets.

A store or category digital twin is a live virtual replica of the category — its SKUs, its shelf, its demand — that a plan can be run against as a system rather than a spreadsheet. It is genuinely emerging: deployed in pockets, promising, not yet routine. Pilot it where the category is large and complex enough to justify the modelling cost.

Technique 5The one to watch, not buy

Agent-based shopper modelling — and its cousin, the LLM-generated synthetic shopper persona of Prompt 12 — is the frontier. The idea is seductive: populate a simulation with thousands of autonomous simulated shoppers, each with their own decision tree and budget, and watch category-level behaviour emerge. Academic frameworks exist for generating synthetic transaction data this way. But the validity is unproven, and the failure mode is specific and dangerous: a synthetic shopper is fluent and confident, and it is tempting to treat its “answer” as data. It is not data. It is a structured hypothesis. Use it to generate questions for real research, never to replace it.

In practiceThe simulation loop

Simulation is not a one-off; it is a loop that wraps the plan before the plan reaches a human:

1. Design the scenario set — Prompt 11 turns a draft plan into a base case, a sensitivity list and a set of named scenarios with explicit kill criteria.
2. Run the scenarios — what-if for the directional question, Monte Carlo for the probabilistic one, virtual shelf testing for the navigability question.
3. Read the results as ranges and probabilities, against the kill criteria.
4. Revise the plan where a scenario breaks it — or halt if a kill criterion fires.
5. Hand the surviving plan, with its outcome ranges attached, to the Governance Agent and then the human approver.

Done this way, simulation is not a fancy add-on. It is the difference between presenting an approver a number and presenting them a bet they can actually price.

Category management was always a discipline with a legal edge. Chapter 3 explained why: the category-captain model asks a self-interested competitor to advise on its rivals’ shelf treatment, and regulators have watched it for two decades. Add AI — faster, more autonomous, harder to interrogate — and every one of those old concerns sharpens. This chapter covers the five risks that matter and the governance that contains them. It is not optional reading. An agentic category system without this chapter is a liability with a user interface.

Risk 1The AI category captain

The traditional category-captain concern is that a leading supplier, advising on the whole category, quietly favours itself — competitive exclusion — or becomes a conduit for coordination. An AI category captain amplifies this in three ways. It operates at a scale and speed no human captain could, so a subtle bias compounds fast. It is harder to interrogate — “the model recommended it” is not an audit trail. And it can ingest and act on more data, including data it should never have touched. The mitigations are the classic ones, made firmer: an AI captain advises, never decides; the retailer retains and exercises final authority; a validator — human or a separately-instructed model — reviews; and the captain’s AI is provably walled off from rivals’ confidential data.

Risk 2Algorithmic price collusion

This is the risk regulators are most actively pursuing, and it has a specific technical root. In a much-cited 2021 simulation, several independent pricing algorithms using reinforcement learning — with no instruction to collude and no communication — spontaneously learned collusive, supra-competitive pricing, simply because mutual restraint was the strategy that maximised each agent’s reward. Regulators have taken the lesson: US enforcers treat algorithmic price-fixing as ordinary price-fixing, and a landmark case over shared pricing software has put the whole field on notice. For category management the rule is blunt: a pricing agent must never be allowed to set prices using rivals’ confidential or signalled data, and never optimise toward an objective that rewards tacit coordination. Prompts 1, 8 and 13 encode exactly this.

Risk 3Hallucination & invented evidence

An LLM’s defining failure mode is the confident, fluent, wrong answer — and in category management the wrong answer wears the costume of a precise number. “Private label is losing 3.2 points of share” reads identically whether it came from the data or from the model’s imagination. Surveys of enterprise LLM use trace a large share of hallucinations to data problems — missing context, training bias. The defences run through this whole guide: the Context Pack’s explicit gap list (Chapter 10), the system prompt’s “NOT IN CONTEXT” rule (Prompt 1), the assumption-and-uncertainty list every prompt demands, and the Governance prompt’s hunt for unsourced numbers (Prompt 13).

Risk 4Data bias & supplier equity

A model trained on biased history reproduces and amplifies the bias. In category management this shows up as systematically uneven output across store types, shopper demographics or regions — and, pointedly, as a structural tilt against smaller suppliers, whose thinner data trail makes them easier for an optimisation to quietly de-range. A plan that cuts small suppliers because they are small, rather than because they fail a fair category test, is both an equity problem and a competition-law exposure. Prompt 13’s equity check exists for this; so does the rule, in Prompt 7, to judge private label and national brands by one evidence standard.

Risk 5Opacity & the un-auditable decision

A recommendation no one can explain is a recommendation no one can defend — to a regulator, to a supplier, to your own board. Research on AI adoption is consistent: people will not, and should not, act on black-box outputs. The answer is explainability by construction: every agent states its reasoning and its evidence (Prompt 1), every plan carries its assumption list, and the whole pipeline keeps an audit trail — what data went in, which agent recommended what, which human approved it, when. If you cannot reconstruct a category decision six months later, you do not have governance; you have luck.

The controlA decision-rights table

The single most important governance instrument is also the simplest: an explicit table of who — or what — may decide what. The principle, from Chapter 9: the more consequential, less reversible and more externally visible an action, the more it demands human approval before execution.

The controlSix guardrails to put in place first

• The Governance Agent runs last and can halt. No plan reaches a human approver without passing Prompt 13; a FAIL stops the plan.
• Confidential-data walls are technical, not promised. A captain’s AI is architecturally prevented from accessing rivals’ confidential data — not merely instructed to ignore it.
• No price coordination, ever. Pricing agents are barred from rivals’ signalled data and from objectives that reward coordination; this is checked, not assumed.
• Every plan carries its assumptions and provenance. No number without a source; no recommendation without its reasoning.
• The full pipeline is logged. Inputs, agent outputs, approvals and timestamps are retained as an audit trail.
• A human owns every decision. The AI is accountable to no regulator; a named person is. Decision rights are written down and honoured.

Everything in Part II converges here. You have a state-of-the-art map (Chapter 8), an architecture (9), a context discipline (10), a prompt library (11), a simulation lab (12) and a governance frame (13). What remains is sequencing — because the organisations in Chapter 8’s 71% did not fail for lack of ambition; they failed by skipping straight to the ambitious part. This chapter gives you a maturity model to locate yourself honestly and a 90-day plan to make the first move count.

Where you standA five-stage maturity model

No single industry-standard maturity model for AI in category management exists; the ladder below adapts general AI-maturity thinking to this discipline. Locate yourself honestly — most organisations are at Stage 1 or 2, and naming that is the start of progress, not an admission of failure.

How to climbCrawl, walk, run

The ladder maps cleanly onto a sequencing rule. Crawl (Stages 1–2): adopt the proven, low-risk wins — AI demand forecasting, computer-vision shelf monitoring — and put copilots in category managers’ hands using the prompt library. Walk (Stage 3): deploy AI assortment and planogram optimisation, and — the real work — unify the data foundation so recommendations can be trusted. Run (Stages 4–5): stand up the agentic operating model, one agent and one category at a time, behind the governance frame. The cardinal error is running before you can walk: agents on a messy data foundation simply automate the production of wrong answers.

The first moveA 90-day plan

Ambition fails on vagueness, so here is a concrete first quarter. It assumes you are at Stage 1 or 2 and want a credible, low-risk, value-proving step toward Stage 3–4. Pick one category — important enough to matter, contained enough to finish — and run this.

Days 1–30 — Foundation

• Choose the pilot category and name a human owner accountable for the pilot.
• Build the Category Context Pack for it — the full Chapter 10 template. This is most of the month’s work and the point of it.
• Audit the data behind the pack: provenance, freshness, gaps. Fix what is cheap to fix; document the rest in section 8.
• Write the rules of engagement — constraints, guardrails, decision rights — and have competition counsel review the antitrust boundaries.
• Define success: the two or three metrics by which the pilot will be judged in day 90.

Days 31–60 — Assisted run

• Run the prompt library end to end on the pilot category — Prompts 1 through 13 — as copilots, with the category manager driving.
• Compare the AI-assisted outputs against how the category was last reviewed manually. Where do they agree? Where does the AI surface something missed — or get something wrong?
• Run the Governance prompt (13) on the resulting plan and act on its flags.
• Capture the friction: every place the model lacked context, every prompt that needed rewording. This feedback improves the pack and the prompts.

Days 61–90 — Simulate, decide, scale-plan

• Run the simulation loop (Chapter 12) on the plan — what-if and Monte Carlo at minimum — and present outcomes as ranges with kill criteria.
• Take the plan through real human approval using the decision-rights table. Execute what is approved.
• Measure against the day-1 success metrics. Be honest; a clear negative result is a real result.
• Write the scale decision: which categories next, which agent from Chapter 9 to stand up first, what the data foundation needs before Stage 4.

The five rulesWhat separates the 29% from the 71%

1. Data before models. The Context Pack is the project. Skip it and you join the 71%.
2. Proven before emerging. Bank the forecasting and copilot wins before piloting agents.
3. One category before the portfolio. Prove it small, learn, then scale — never the reverse.
4. Governance from day one. The decision-rights table and the Governance Agent are not a later add-on.
5. A human owns it. Every plan, every decision, every quarter — named, accountable, in the loop.