# The Retail AI Vendor Map: A Working View

> Second edition, verified 1 October 2026. Three layers of the retail AI market, revised category judgements, and why US checkout moved back inside the assistant in September.

Source: https://barkingcat.ai/thinking/retail-ai-vendor-map-october-2026/

## Thinking · October 2026

**Second edition, verified 1 October 2026.** The July 2026 edition remains published, unchanged, at [barkingcat.ai/thinking/retail-ai-vendor-map/](https://barkingcat.ai/thinking/retail-ai-vendor-map/). A September revision was drafted and overtaken by events before it could be published, which says something about the pace of the agentic layer. The practice writes down what it believes, when it believed it, and lets the record keep score. What was retained, revised, retired and added since July is listed at the end of this page.

The market data still explains the mood in every boardroom. McKinsey’s 2026 State of AI survey (1,719 respondents across 97 countries, published in August) finds nearly nine in ten organisations using AI regularly in at least one function and 44% reporting that AI is scaling across the enterprise. Yet only 37% attribute any EBIT impact to it, essentially unchanged on 2025, and the share qualifying as high performers, attributing at least 5% of EBIT to AI and describing the value as significant, is 6%, also unchanged. NVIDIA’s 2026 retail survey has nine in ten retailers increasing AI budgets this year, half of them by 10% or more (a chip vendor’s survey, so directional). BCG’s 10-20-70 rule, its heuristic that 70% of the effort and the difficulty in AI sits in people and process rather than technology, still holds and is still ignored at budget time. Spending is accelerating faster than results, which is precisely the condition under which vendor selection matters more than vendor marketing suggests. Most vendor confusion starts with a simpler mistake: not knowing which layer of the market you are buying from.

There are still three. What changed since July is the boundary between them.

## 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, revised since July.

First, the supply chain is regulated, and July’s caution has become the record. On 12 June the US Commerce Department imposed export controls on Anthropic’s two newest models, Fable 5 and Mythos 5. Unable to verify nationality in real time across its user base, Anthropic suspended both for every customer worldwide. The controls were lifted on 30 June and general access to Fable 5 returned on 1 July; Mythos 5 came back only for a limited set of US organisations. The pattern has since repeated by design rather than by order. Anthropic’s Mythos 5.1, released on 1 September, is available only to vetted US organisations, and Google’s Gemini 4 Argon, released on 30 September, went to cyber-security partners only. The most capable models now reach a restricted group first and the general market later, if at all. Multi-model architecture is the default recommendation, and any application vendor that cannot tell you which model sits underneath its product, and what happens when that model goes dark, has not finished its own risk assessment.

Second, the economics keep moving mid-business-case. September alone brought GPT-6 Astra, GPT-6 Sol and Luna (the latter two priced at half the GPT-5.6 series), Anthropic’s Opus 5.5 and Sonnet 5.5, and a 75% cut to the price of cached input on Anthropic’s Fable 5.1. Cached input that cost roughly a tenth of the standard rate in July now costs a fortieth to a twentieth on the newest models. In the same McKinsey survey that reports the flat ROI, 20% of respondents cite AI operating costs as a constraint on further use. Cheaper per token and dearer in aggregate is a combination most business cases have not been built to hold.

Third, the boundary with Layer 2 is dissolving from both sides. Fin, the customer-service agent Salesforce has now bought, launched on GPT-4, later leaned on Claude, then post-trained its own model, Apex, to reduce dependence on the frontier labs. On 15 September Salesforce itself announced Koa, its own reasoning model for CRM work, post-trained on NVIDIA’s open Nemotron model, weeks after making Claude a default inside its platform. Application vendors building their own models while frontier labs build applications is the pattern to watch, because it changes who you are actually contracting with. The commercial differences that matter remain distribution and ecosystem; benchmark decimal places do not. No retailer should be building strategy on which lab tops a leaderboard this month, and this month there have been several.

## Layer 2: The applications

The products retailers actually buy, mapped to the retail value chain. The honest view by category, updated where the evidence moved.

### Search and product discovery

Still the fastest-payback AI investment on a retail site, and still the least glamorous. Two traps, both larger than in July. 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 discovery keeps migrating off-site. Adobe’s August report, covering US retail to July, has AI-referred visitors converting 60% better than other traffic (up from 54% in May) and generating 53% more revenue per visit; its holiday forecast, published on 28 September, expects AI-referred traffic to US retail sites to rise 130% year on year across November and December. Shopify reported AI-driven sessions and orders each tripling in its second quarter. All vendors with something to sell, so directional, and the denominator remains small. Contentsquare’s benchmark of 99 billion sessions put AI-referred visits at 0.2% of the total in late 2025, and IMRG’s UK panel of 168 retailers had assistants at 0.39% of traffic, converting at 2.28% across the 120 retailers that reported it. The share is small and the conversion is high, and the two numbers are routinely confused, which is why the practice published [a separate piece on exactly that confusion](https://barkingcat.ai/thinking/two-numbers-ai-visibility/) in August. When discovery happens inside an assistant, the on-site search engine is no longer the front door. It is the room the front door opens into, and it had better be tidy.

### Personalisation and experience

Squeezed from both directions, and the squeeze now has a number. McKinsey’s 2026 survey found 32% of organisations had decided against buying at least one software product or feature because they could build it internally with agentic coding tools; among high performers the figure is nearly half, against 31% of everyone else. For consumer goods and retail respondents it is 29%. What needed a specialist vendor in 2023 is increasingly a well-prompted model plus a customer data platform, and increasingly something the retailer’s own team assembled. The durable vendors are the ones who own the experimentation discipline and the unified data layer rather than 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 remains 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.

July’s figure for how bad that data is (20 to 30% of SKUs with missing or wrong attributes, attributed to search-vendor onboarding audits) has been retired. It could not be traced to a named primary source, and this page does not carry numbers it cannot trace. Two verifiable figures replace it. Adobe’s April benchmark scored US retail product pages at 66% machine-readable on average against 75% for homepages and 74% for category pages, with the best-performing sites at 82.5% and the worst at 54.2%: the pages that convert are the pages an agent reads worst. And the practice’s own [Commerce Experience Index](https://barkingcat.ai/thinking/commercial-experience-index-edition-two/), published in August, found 16 of 20 Premier League club shops returning a 404 at the first place an AI agent looks for product data.

### Customer service

The most mature category, with the clearest unit economics, and the one that consolidated most expensively this year. Salesforce completed its acquisition of Fin, formerly Intercom, on 10 September, around three months ahead of the timetable it gave when the roughly $3.6 billion deal was signed in June. Fin’s 30,000-plus customers now sit inside Salesforce. Acquisition risk in vendor selection is a matter of record rather than caution.

Pricing has converged on per-resolution outcomes, and the published rates sit in a band that did not move at the close: HubSpot at $0.50 per resolved conversation since April, Fin at $0.99 per outcome, Zendesk at $1.50 committed and $2.00 pay-as-you-go, Salesforce Agentforce at $2.00 per conversation whether or not it resolved anything. Sierra, Decagon and Ada still do not publish. The business case writes itself only if resolution is measured honestly, so read each vendor’s definition of a billable outcome before comparing rates; Fin’s lists five separate conditions. The KPI shift from containment to resolution remains the tell for whether a deployment is real. Woolworths’ Olive, in its deterministic form, was already handling more than 70% of contact-centre interactions before its agentic rebuild. The interesting number in that story is the eight in-house “judges” Woolworths built to check every agentic answer, recalculating the arithmetic and verifying the legal and food-safety claims, before a customer sees it. That is what honest measurement looks like in production, and the retailer built it in-house.

### Pricing and markdown

High value, high sensitivity, and the category where governance now bites in statute as well as in headlines. The distinction regulators are drawing is between dynamic pricing, which responds to market conditions such as stock, demand and season, and personalised or surveillance pricing, which responds to the individual. Consumer Reports’ December test found around 74% of Instacart items offered at multiple prices simultaneously, with some shoppers paying up to 23% more for the identical item at the same store; Instacart halted the experiments within days. Maryland’s ban, covering food retailers and delivery platforms, and Connecticut’s limits both took effect today, 1 October. New Jersey signed its Fair Price Protection Act on 23 July, covering groceries and other necessities. New York’s One Fair Price Act passed the legislature in June and still awaits the governor; California’s equivalent died when the session ended on 31 August. The FTC proposed an enforcement policy in August that would treat undisclosed personalised pricing as potentially deceptive, with comments closing on 25 September. In the EU, the AI Act’s Article 50 transparency duties took effect on 2 August, with fines of up to 15 million euros or 3% of worldwide turnover. Guardrails before algorithms, and now guardrails before the algorithm becomes a court exhibit.

### Supply chain and planning

The deepest ROI in retail AI and the longest, most expensive road to it. The two deflationary numbers hold, with their dates attached this time. Sage’s 2026 State of Supply Chain report, a survey of more than 200 retail and wholesale operators by a supply-chain software vendor and therefore directional, found only 10% with AI live in supply-chain workflows, and adoption tracking data readiness and visibility rather than interest. Gartner’s prediction that more than 40% of agentic AI projects will be cancelled by the end of 2027 dates from June 2025 and has been recirculated through 2026 as though new; Gartner’s April 2026 Hype Cycle places agentic AI at the peak of inflated expectations, with 17% of organisations deployed and more than 60% expecting to deploy within two years. McKinsey’s respondents, meanwhile, report supply chain as the function where AI most often reduces cost. Both things are true. The value is real and the road is long. Every vendor in the category is selling agents; the honest question is whether your master data and planning discipline can feed one.

### Agentic commerce

Still the newest column, and the one where this edition revises the July judgement. In July the position was that consumer discovery through assistants had gone mainstream, autonomous purchase had not, and the first attempt at buying inside the chat had failed. OpenAI retired Instant Checkout in March after a short life with few merchants live, and Walmart withdrew after in-chat checkout converted at one-third the rate of click-through. The market settled on discover in the assistant, transact with the merchant.

September reopened the question in the United States, by default rather than by demand. Meta launched Muse, a personal shopping agent, on 8 September, and Shop Pay checkout inside it followed on 22 September. Shopify switched agentic checkout on by default for eligible US merchants on 21 September, opt-out by channel, and a week later added checkout tools for browser-based agents. Around the same date Google switched on native checkout in AI Mode and Gemini for eligible US Shopify merchants, again opt-out. Amazon, which replaced Rufus with Alexa for Shopping in May and lets it buy on third-party sites, blocked Muse from its own. In each new arrangement the retailer remains merchant of record, so it still holds the transaction and the customer data, but the purchase moment has moved back into the assistant, and the merchant’s consent has moved from opt-in to opt-out.

The revised judgement, for a UK board: the second attempt at in-chat buying is better designed than the first, because it leaves the retailer as merchant of record, and it arrived through platform settings rather than through retailer strategy. There is still no UK date for Google’s Universal Cart; Google’s wording remains “later to the U.K.” Flagship Advisory Partners’ tracker of 80 large US and EU merchants, last updated on 31 August, found only four supporting agent-led checkout, all in the US, with AI traffic still under 1% of retail visits. A UK retailer on Shopify should know today which of those settings will apply to it when they cross the Atlantic, and who in the business decides.

The merchant-side agent is the pattern that has shipped most cleanly outside the US: Woolworths’ Olive reached Australian customers in July, reading a photographed recipe, applying rewards and building the basket, with the retailer keeping the entire relationship.

The demand evidence is real and vendor-sourced (Adobe and Shopify, above). The caution is real too, and only partly independent. Bain’s November 2025 consumer research found shoppers trust a retailer’s own on-site agent around three times more than a third-party AI platform, which only about 7% would trust to shop for them. Checkout.com’s June 2026 research found 41% of Britons trusting no organisation to run a shopping agent for them, and an ACI Worldwide survey of more than 2,000 UK adults by YouGov, published the same month, found 60% would stop using one after a single mistake. Both companies sell payments infrastructure, so both have an interest in trust being the problem; the direction is still worth having. Narvar’s holiday survey has 65% of consumers planning to use AI for some part of their holiday shopping against 8% of retailers very confident in their ability to serve them; a returns vendor, so directional, but the gap is the story. Infrastructure is still racing ahead of consent. The readiness question is still boring and still decisive: product data quality, and since 2 August, a disclosure the customer can see.

## The governance strand that arrived in August

In July, governance was a horizon item. On 2 August it became an obligation. Article 50 of the EU AI Act now requires providers and deployers to tell people when they are interacting with an AI system unless it is obvious, and to mark AI-generated content so it can be detected. The duties attach to any AI system whose outputs are used within the European Union regardless of where the provider or the retailer sits, so a UK retailer trading into Ireland or France is in scope. Generative systems already on the market have until 2 December to meet the machine-readable marking requirement; systems placed on the market after 2 August get no grace period. The Commission published its guidelines on 20 July and confirmed the code of practice on AI-generated content as adequate for demonstrating compliance; signing it is voluntary, following it is the easiest way to evidence the duty. The heavier high-risk obligations were deferred to December 2027 and August 2028, which is why some retailers have read 2 August as a quiet date. It was not. No enforcement action had been reported by 1 October, which is the usual silence before the first one.

Three consequences for the map. Every Layer 2 category with a customer-facing surface, which is service, discovery and agentic commerce, now carries a disclosure obligation the customer can see. Pricing is a governance category on both sides of the Atlantic at once. And the part of Layer 3 that survives commoditisation, governance and sequencing, now has a statute behind it. The operator question is short: who in your organisation owns AI governance, by name, with a budget? In most retailers the answer is a working group, which is to say nobody.

## Layer 3: The delivery market

Where implementation lives, and where commodity work dies, faster than it was dying in July. The layer runs from global strategy houses through digital consultancies and systems integrators to boutique specialists and a still-growing cohort of certificate-course consultants. Two structural facts now govern it. Implementation of commodity tooling is commoditising toward zero as vendors absorb it into product; Woolworths took Olive from a deterministic bot to a customer-facing agent in around six months on a platform product, with the governance layer the only part built by hand. And the buyer is becoming the builder: the 32% of McKinsey’s respondents who built rather than bought are the same clients who no longer need an integrator for the middle of the work. Gartner added a sharper edge on 29 September, predicting that 70% of enterprises will abandon agentic AI built for them by vendors’ forward-deployed engineers by 2028. An analyst forecast, not data, but it points at the same question. 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, and McKinsey’s flat 6% suggests the half has not shrunk in a year. The map is for sequencing, and the sequence starts with the unglamorous questions: what does your product data look like, what can your operating model absorb, who in the room does not earn from the answer, who signs the disclosure, and, since September, which of your platform’s default settings has already answered the agentic question for you.

## What changed since July

**Retained.** The three-layer structure. The category judgements on search, personalisation, content, supply chain and delivery. BCG’s 10-20-70. Gartner’s 40% cancellation forecast, now dated to June 2025. Sage’s 10%. The 41% and 60% consumer trust figures, now correctly attributed to Checkout.com and ACI Worldwide. The Walmart one-third conversion figure. IMRG’s 0.39%.

**Revised.** The budget figure: July cited 97% of retailers increasing AI budgets, which traces to NVIDIA’s 2025 survey; the 2026 survey says nine in ten. The ROI figures: July’s 31% is replaced by McKinsey’s 2026 figure of 37% attributing any EBIT impact, with the 6% high-performer share unchanged. The export-control episode, now past tense, with Mythos access still restricted. EU transparency duties, now in force. Woolworths’ Olive, now live to customers since July. Adobe’s traffic and conversion data, refreshed to July. Customer service, now with a completed $3.6 billion acquisition and a published rate band. The agentic commerce judgement: discover in the assistant and transact with the merchant held from March to August; in September, US checkout moved back into the assistant by platform default, with the merchant kept as merchant of record.

**Retired.** The 20 to 30% missing-attribute figure, untraceable to a primary source. Replaced by Adobe’s machine-readability benchmark and the practice’s own Commerce Experience Index finding.

**Added.** McKinsey’s build-versus-buy figures. The Fin and Apex example of an application vendor building its own model, and Salesforce’s Koa. Restricted first release as the new pattern at the model frontier. Published per-resolution pricing. US surveillance-pricing statutes and the FTC’s proposed policy. The governance section. Meta’s Muse, Shopify’s and Google’s default-on agentic checkout, and Amazon’s block. Bain’s trust ratio. Flagship Advisory Partners’ August tracker. Gartner’s forward-deployed engineering forecast.

Landscape verified 1 October 2026. The July 2026 edition remains published, unchanged. The model layer moves weekly and the agentic layer monthly; this edition is the proof.

## Sources
- McKinsey, The state of AI in 2026: On the road to ROI, August 2026 (1,719 respondents, 97 countries; survey 4 May to 8 June 2026).
- NVIDIA, State of AI in Retail and CPG 2026, 7 January 2026 (vendor survey, directional).
- BCG, Where’s the Value in AI?, 24 October 2024 (10-20-70 heuristic; BCG sells AI transformation consulting).
- US Commerce Department controls of 12 June 2026 and their lifting on 30 June 2026, as reported by BleepingComputer (30 June 2026), Al Jazeera (1 July 2026) and Medianama (July 2026); Anthropic, Claude Fable and Mythos 5.1, 1 September 2026; Google Gemini 4 Argon as reported by TechCrunch, 30 September 2026.
- OpenAI, GPT-5.6 (9 July 2026) and GPT-6 Astra (September 2026) announcements; TechCrunch, 22 September 2026 (GPT-6 Sol and Luna); Anthropic published API pricing, 1 October 2026.
- Salesforce, Salesforce Signs Definitive Agreement to Acquire Fin, 15 June 2026; Salesforce Completes Acquisition of Fin, 10 September 2026; Salesforce, Koa announcement, 15 September 2026.
- Adobe, AI Traffic Trends Report, 19 August 2026; Adobe holiday forecast press release, 28 September 2026; Adobe blog, 16 April 2026 (machine-readability benchmark). Vendor data, directional.
- Shopify Q2 2026 results, 5 August 2026; Contentsquare press release, 29 January 2026 (99 billion sessions, Q4 2025). Vendor data, directional.
- IMRG, Tech and AI Connect, July 2026 (traffic share, 168 retailers; conversion, 120 retailers).
- Barking Cat, Commerce Experience Index Edition Two, August 2026.
- Published vendor pricing pages, checked 1 October 2026: HubSpot, Fin, Zendesk, Salesforce Agentforce (definitions vary by vendor).
- iTnews, 28 April and 29 June 2026; iTWire, 30 June 2026 (Woolworths Olive, agentic judges, contact-centre share).
- Consumer Reports, December 2025 (Instacart pricing test); Maryland HB 895 (2026); Connecticut Attorney General, 16 September 2026; Office of the Governor of New Jersey, 23 July 2026; New York Senate, A.9349B; KALW, 1 September 2026 (California AB 2564); FTC, 3 September 2026, and Morgan Lewis, August 2026 (proposed personalised pricing policy).
- European Commission, guidelines on Article 50 transparency obligations, 20 July 2026 (page updated 6 August 2026); Faegre Drinker, July 2026; Baker Botts, 10 September 2026.
- Sage, 2026 State of Supply Chain Report, February 2026 (vendor survey, directional).
- Gartner, press releases of 25 June 2025 and 29 September 2026; Gartner, Hype Cycle for Agentic AI, 15 April 2026.
- CNBC, 20 and 24 March 2026; Forbes, 10 March 2026 (Instant Checkout withdrawal); WIRED, 18 March 2026 (Walmart).
- Meta, Introducing Muse, 8 September 2026; The Paypers, 22 September 2026 (Shop Pay in Muse); Yahoo Finance and Forkast, 24 September 2026 (Shopify default-on agentic checkout); TechCrunch, 28 September 2026 (Shopify browser-agent checkout); Google Shopping blog, 16 September 2026; PPC Land, September 2026 (AI Mode checkout for Shopify merchants); Payments Dive, 23 September 2026 (Amazon blocks Muse); Amazon, Alexa for Shopping, 13 May 2026.
- Flagship Advisory Partners, Market Tracker: Agentic Commerce, August 2026 (published 31 August 2026).
- Bain and Company, press release of 13 November 2025.
- Narvar with eTail Insights, 2026 Holiday Shopping Report, August 2026 (vendor survey, directional).
- Checkout.com, agentic commerce consumer research, 11 June 2026 (vendor survey: Checkout.com sells payments processing), as reported by InternetRetailing, 11 June 2026.
- ACI Worldwide with YouGov, UK consumer survey (more than 2,000 adults), 30 June 2026 (vendor survey: ACI sells payments technology), as reported by Retail Technology Innovation Hub, 30 June 2026.

## Contact

steve@barkingcat.ai

Barking Cat Advisory Limited. Registered in England and Wales, company number 17338003.