The market data explains the mood in every boardroom: 97% of retailers plan to increase AI budgets, yet McKinsey finds only 6% of organisations generating enterprise-level value, only 31% report positive ROI, and BCG places 70% of AI value in people, process and operating model rather than technology. The constraint is organisational and commercial, not technical. Most vendor confusion starts with a simpler mistake: not knowing which layer of the market you are buying from.
There are three.
Layer 1: The models
The frontier providers: Anthropic, OpenAI, Google, Meta, xAI, Mistral, and the open-weights ecosystem. Retailers rarely buy these directly below enterprise scale; they consume them through the application layer or hyperscaler contracts.
Three things a board needs to know about this layer, and only three. First, it is now a regulated supply chain: the June and July 2026 export-control episode, in which frontier releases were reviewed, restricted and in one case temporarily suspended by government directive, made single-model dependency a supply-chain risk in the literal sense. Multi-model architecture is now the default recommendation, not a sophistication. Second, token prices are collapsing, and prompt caching takes cached input to roughly a tenth of standard rate, which changes the economics of most use cases mid-business-case. Third, the differences that matter commercially are distribution and ecosystem, not benchmark decimal places. No retailer should be building strategy on which lab tops a leaderboard this month.
Layer 2: The applications
The products retailers actually buy, mapped to the retail value chain. The honest view by category:
Search and product discovery
Consistently the fastest-payback AI investment on a retail site, and the least glamorous. Two traps. Every vendor in the category is only as good as the catalogue data underneath it, which is why data readiness questions come before any vendor conversation. And agentic commerce is quietly commoditising the category from above: when discovery happens inside an assistant, the on-site search engine is no longer the front door.
Personalisation and experience
Being squeezed from both directions: platform-native features below, frontier-model capability above. What needed a specialist vendor in 2023 is increasingly a well-prompted model plus a customer data platform. The durable vendors are the ones who own the experimentation discipline and the unified data layer, not the recommendation widget. The advice pattern that holds: fix identity resolution before buying personalisation, because personalising on fragmented data degrades whichever platform you choose.
Content and catalogue
Content generation is the most commoditised category, the easiest pilot, and the least defensible vendor moat; frontier models eat upward into it monthly. The durable value sits in the unglamorous catalogue layer, because agent-mediated commerce is fed by product data, and most retailers’ product data is not fit for it. Search-vendor onboarding audits commonly find 20 to 30% of SKUs with missing or wrong attributes.
Customer service
The most mature category, with the clearest unit economics: pricing has converged on per-resolution outcomes, which means the business case writes itself if, and only if, resolution is measured honestly rather than as deflection. The KPI shift from containment to outcomes is the tell for whether a deployment is real. The category is also consolidating fast, so vendor selection now has to weight acquisition risk.
Pricing and markdown
High value, high sensitivity, and the category where governance bites first: dynamic pricing is where AI ambition meets consumer-fairness headlines and, in the EU, transparency duties arriving in August 2026. Guardrails before algorithms.
Supply chain and planning
The deepest ROI in retail AI and the longest, most expensive road to it. Two numbers deflate the vendor decks: Sage’s 2026 State of Supply Chain report found only around 10% of operators with AI actually live in supply chain workflows, and Gartner expects more than 40% of agentic AI projects scrapped by the end of 2027 on cost, unclear value or weak risk controls. Every vendor in the category is selling agents; the honest question is whether your master data and planning discipline can feed one.
Agentic commerce
The newest column and the noisiest. What is verified as of mid-July 2026: consumer discovery through assistants is mainstream, autonomous purchase is not, and the first version of buying inside the chat has already failed commercially. The market has pivoted to discover in the assistant, transact with the merchant, which is better news for retailers than the headlines suggest, because the merchant keeps the customer, the login and the loyalty data. The merchant-side agent is the pattern to watch: Woolworths’ Olive, rebuilt on Google’s enterprise platform and rolled out to customers this month, reads a photographed recipe, applies rewards and builds the basket, with the retailer keeping the entire relationship. The demand evidence is real (AI-referred traffic to US retailers grew 393% year on year in Q1 2026 and now converts materially better than average, per Adobe), and so is the caution: 41% of Britons trust no organisation to run a shopping agent for them, 60% would stop using one after a single mistake, and consumers trust agents to find a price far more than to fix a problem (practice research, July 2026). Infrastructure is racing ahead of consent. The readiness question is boring and decisive: product data quality.
Layer 3: The delivery market
Where implementation lives, and where commodity work dies. The layer runs from global strategy houses through digital consultancies and systems integrators to boutique specialists and a fast-growing cohort of certificate-course consultants. The structural fact that matters: implementation of commodity tooling is commoditising toward zero as vendors absorb it into product. One-line integrations, template agents and platform-native features keep swallowing what was billable work a year ago. What survives is judgement: sequencing, governance, and the willingness to say you do not need this. When buying delivery, the question is not the day rate; it is whether the adviser earns from the build they are recommending.
The diagnosis underneath the map
A business that buys layer 2 software and expects layer 1 magic without layer 3 change is asking an institution to behave against its nature. Half the vendor conversations in retail right now are exactly that. The map is not a shopping list; it is a sequencing tool, and the sequence starts with the unglamorous questions: what does your product data look like, what can your operating model absorb, and who in the room does not earn from the answer.
Landscape verified 14 July 2026. The model layer moves weekly and the agentic layer monthly; figures are dated for a reason.
Back to Thinking