Something quietly significant happened when Anthropic released the Model Context Protocol (MCP) and it began attracting rapid adoption across the AI ecosystem. For the first time, there is a credible, emerging standard for how AI agents — not humans — interact with external services. That shift matters enormously for ecommerce. The question is no longer whether AI will influence purchasing decisions; it is whether your website is architecturally capable of being transacted with by a non-human buyer acting autonomously on a customer's behalf.
Agentic commerce — the idea that software agents will browse, evaluate, and complete purchases without direct human input — is moving from theoretical to operational faster than most retailers have noticed. UK organisations that instrument their product catalogues, pricing logic, and checkout flows for agent-readable interaction now stand to capture a meaningful first-mover advantage. Those that wait risk being structurally invisible to the next wave of AI-driven purchasing behaviour.
What MCP Actually Changes for Ecommerce
Anthropic's Model Context Protocol provides a standardised way for large language models to call external tools and services in a structured, authenticated, and stateful manner. Think of it as an API layer specifically designed for AI agents rather than human developers. Where a traditional REST API was built with a developer's mental model in mind, an MCP endpoint is designed to be discovered and used by a reasoning model that has no prior knowledge of your system — it reads the schema, understands the intent, and acts accordingly.
For ecommerce specifically, this means an AI agent managing a procurement workflow or acting on a consumer's shopping preferences can, in principle, query your product catalogue for availability, check pricing against a budget constraint, apply a discount code, and complete a checkout — all without a human touching a browser. The prerequisite is that your site exposes the right structured endpoints and accompanying metadata that agents can reliably interpret. Without that instrumentation, your site simply does not exist to the agent, regardless of how well it ranks in traditional search.
The Role of llms.txt in Agent Discovery
Alongside MCP, a lightweight but important convention has emerged: the llms.txt file. Analogous to robots.txt for search engine crawlers, an llms.txt file placed at the root of your domain gives AI systems a structured, markdown-formatted overview of your site's purpose, key resources, and how content should be interpreted. Early adopters in the developer tools and SaaS space have already implemented it; ecommerce has been slower to follow.
For a retailer, a well-constructed llms.txt is not merely a signal of technical awareness — it is a practical navigation layer. It can direct an AI agent to your product API, explain the structure of your catalogue taxonomy, clarify your returns policy in plain terms, and indicate whether guest checkout is supported. These are precisely the facts an autonomous purchasing agent needs to resolve before committing to a transaction. Getting this file right requires collaboration between your content, commercial, and engineering teams, but the implementation effort is modest relative to the strategic value it unlocks.
Instrumenting Your Catalogue, Pricing, and Checkout
Preparing for agentic commerce is fundamentally an exercise in structured data and API design. Start with your product catalogue: every SKU should be accessible via a queryable endpoint that returns machine-readable attributes — dimensions, materials, compatibility, lead times, and stock levels — not just marketing copy written for humans. Schema.org Product markup remains valuable for traditional SEO, but for agent interaction you need something richer and more dynamic: an endpoint that can respond to queries like 'show me all variants of this product under £150 with next-day delivery available'.
Pricing logic presents its own complexity. Many UK retailers have promotional rules, volume discounts, trade pricing tiers, and currency considerations baked into systems that were never designed to be interrogated programmatically in real time. Exposing a pricing API that can accurately reflect the price a specific user would pay — including VAT, delivery, and active promotions — is a non-trivial engineering task, but it is the critical step that separates a site that can be browsed by an agent from one that can actually be transacted with. Checkout flows, similarly, need to support authenticated, headless completion paths. If your checkout requires JavaScript rendering, cookie-based session state that breaks outside a browser, or CAPTCHAs at the payment step, you have structural barriers that will prevent agent-driven transactions entirely.
B2B vs B2C: Where the Urgency Differs
The agentic commerce opportunity is not uniform across all retail models. For B2B ecommerce — procurement, wholesale, trade supply — the case for urgency is most acute. Procurement agents are already being deployed by larger enterprises to automate supplier research and purchasing, and the organisations running those agents will naturally route spend towards suppliers whose catalogues are machine-readable and whose checkout flows support programmatic completion. If you supply to businesses, your portal's technical architecture has become a commercial capability question.
In B2C, the timeline is somewhat longer, but the trajectory is clear. Consumer AI assistants capable of completing purchases are already in limited deployment, and the friction of managing wishlists, comparing prices across retailers, and remembering preferred configurations is exactly the kind of task consumers will increasingly delegate. Retailers in considered-purchase categories — electronics, furniture, specialist outdoor equipment — are likely to see agentic behaviour emerge earlier than those in impulse-led categories. Understanding where your product category sits in that curve should inform how aggressively you prioritise the technical work.
The practical starting point for most UK retailers is an honest audit of what a non-human caller can currently learn from and do with your site. Can it discover your product range programmatically? Can it determine the true price for a given customer type? Can it complete a transaction without browser-based interaction? For the majority, the answer to at least one of those questions will be no — and that gap is the work.
At iCentric, we help organisations map the distance between their current technical architecture and the requirements of agent-ready commerce, then build the endpoints, documentation, and integration patterns that close it. The window in which this work translates into genuine competitive differentiation is real but finite. The retailers who move now will be the ones whose catalogues are already indexed, whose checkout flows are already validated, and whose pricing APIs are already trusted by the agents that begin routing meaningful transaction volume in the months ahead. The cost of acting early is modest. The cost of acting late, in a channel where discoverability is binary, is considerably higher.
What is Model Context Protocol (MCP) and who developed it?
Model Context Protocol is an open standard developed by Anthropic that defines how AI language models communicate with external tools and services in a structured, authenticated way. It provides a consistent interface so that AI agents can discover and interact with APIs without needing bespoke integration for each service. It has gained rapid adoption across the AI ecosystem since its release.
Do I need to rebuild my entire ecommerce platform to support AI agents?
Not necessarily. Many retailers can add agent-readiness incrementally by exposing structured API endpoints alongside their existing platform, implementing an llms.txt file, and reviewing their checkout flow for headless compatibility. A thorough technical audit will reveal which gaps are cosmetic and which require deeper architectural changes. Full platform rebuilds are rarely the right starting point.
How does llms.txt differ from robots.txt?
Robots.txt instructs search engine crawlers which pages to index or ignore. An llms.txt file, by contrast, is a structured markdown document that gives AI language models a contextual overview of your site — its purpose, key resources, API locations, and how content should be understood. It is designed for comprehension and navigation by a reasoning model, not just crawl permission control.
Which ecommerce platforms currently support MCP endpoints out of the box?
As of now, native MCP support is not widespread in mainstream ecommerce platforms such as Shopify, Magento, or WooCommerce. Most implementations require custom development to expose compliant endpoints. However, the ecosystem is evolving quickly and some headless commerce platforms and API-first solutions are beginning to add MCP-compatible interfaces as demand increases.
How do AI agents handle authentication and payment security during autonomous transactions?
This is an actively evolving area. Current approaches typically involve the agent using OAuth tokens or API keys issued by the user to authenticate on their behalf, with payment credentials stored in a secure wallet or delegated payment service. Retailers need to ensure their checkout APIs support token-based authentication and do not require browser-rendered payment flows that break outside a standard HTTP client.
What are the legal and liability implications of AI agents making purchases on behalf of consumers?
UK consumer contract law generally holds that a contract is formed when an offer is accepted, regardless of whether a human or software agent performs the action — provided the human has authorised the agent to act. Retailers should review their terms and conditions to ensure they address agent-initiated transactions, and should consider how disputes, returns, and cancellations will be handled when no human was directly present at the point of purchase.
How should product descriptions be written differently for AI agents compared to human shoppers?
AI agents prioritise factual, structured, and unambiguous information over persuasive language. Attributes like exact dimensions, material specifications, compatibility details, weight, and delivery lead times are more useful to an agent than lifestyle copy or brand narrative. Ideally, these attributes are exposed as structured data fields in your API rather than embedded in free-text descriptions, though clear plain-English descriptions in llms.txt and product schema remain valuable as supplementary context.
Is there a risk that competitors could use AI agents to scrape pricing data from my MCP endpoints?
This is a genuine consideration. MCP endpoints should be protected with appropriate authentication so that only authorised agents — typically those acting on behalf of verified users — can access sensitive pricing data. Rate limiting, scope-restricted API keys, and monitoring for unusual query patterns are standard mitigations. The risk is manageable and broadly analogous to the existing risk of competitor price monitoring via traditional web scraping.
How will agentic commerce affect customer relationship management and loyalty programmes?
When an agent completes a purchase on a customer's behalf, the human touchpoints that traditionally build brand loyalty — browsing, discovering, being delighted by a UI — are bypassed. Retailers will need to rethink how loyalty is built and maintained in an agent-mediated relationship. This may shift emphasis towards post-purchase experience, personalised agent instructions configured by the customer, and loyalty APIs that agents can query and apply programmatically.
What internal teams should be involved in preparing a site for agentic commerce?
Effective preparation requires collaboration across engineering (API design and MCP endpoint implementation), ecommerce and merchandising (catalogue data quality and taxonomy structure), commercial teams (pricing logic and promotional rules), legal (terms of service and consumer law compliance), and security (authentication design and access controls). Treating it as a pure engineering task without commercial and legal input is one of the most common mistakes organisations make when starting this work.
Get in touch today
Book a call at a time to suit you, or fill out our enquiry form or get in touch using the contact details below