Agentic commerce, beyond the buzzword

An AI agent is a system able to carry out a sequence of autonomous actions to reach a goal, taking intermediate decisions without human sign-off at every step. It is not a chatbot. It is not an “AI-powered” product recommendation. It is a system that perceives a context, plans a sequence, calls on tools and adjusts its course according to the results it obtains.

Applied to digital commerce: an agent is given a loose mandate — “a present for my 8-year-old nephew who likes Lego, budget €60” — browses a catalogue, checks availability in real time, compares options, puts forward a recommendation and prepares the basket. Without the buyer clicking anywhere between their intention and the final proposal.

THE DEFINITION

An agentic commerce system runs a purchase journey autonomously, from the expression of a need through to the recommendation or the assembled basket, deciding at each step within a scope set by the user. Final approval and the payment trigger can remain human: that is the Human in the Loop model, the most widespread form today.

What sets agentic systems apart from conventional AI in e-commerce is not the sophistication of the language model (LLM). It is the ability to act: read, reason, act, check, start over if needed. The model is the brain; the hands are the tools you give it — querying your catalogue, checking stock in real time, obtaining an up-to-date price, assembling a basket…

A year ago the working assumption was simple: every brand would deploy its own sales agent, on its own site, and its customers would come and use it. That assumption is out of date. What is emerging instead are the Super Agents: general-purpose AI assistants that hundreds of millions of users have already adopted — ChatGPT (OpenAI), Gemini (Google), Claude (Anthropic), and soon Siri (Apple, rollout under way). These platforms are not chatbots. They are systems able to plan, to equip themselves with tools and to act, including to buy. In July 2025, year on year, e-commerce traffic coming from these generative AIs jumped by 4,700% (Adobe 2025). It is also starting to take other forms, with the Buy Button on Gemini or the Business Agent in Google’s AI Mode. Not yet rolled out in Europe, but the trajectory is global: United States, Canada, India, Japan, Australia…

The dominant model will not be one agent per merchant site, but a global agentic channel — and that channel is the AI assistants already in your customers’ pockets. Your challenge is not to build one. It is to be findable, interpretable and actionable by the ones that already exist. This logic does not rule out site-specific agents: they remain relevant for particular use cases — making internal search easier, or handling substitutions on complex B2B baskets… The two models will coexist. But the strategic priority is to be readable and actionable by the general-purpose agents already installed with your customers.

Your customers’ AI assistant does not get inside your systems: it uses what you put within its reach. For it to interact with your catalogue, your stock and your order system, you need a standardised communication layer. Three protocols structure this channel in 2026:

• UCP (Universal Commerce Protocol): an open standard initiated by Google and backed by e-commerce players (Walmart, Nike, Zalando, Etsy, Visa, Mastercard…). It aims to standardise the whole journey, from discovery to basket, checkout and payment, without the agent having to know each retailer’s systems. Its flagship use today: agentic checkout in AI Mode and Gemini.
• ACP (Agentic Commerce Protocol): launched by OpenAI, it acts as a universal binding layer between merchants’ catalogues and LLMs.
• WebMCP (Web Model Context Protocol): a web standard co-developed by Google and Microsoft that lets sites expose tools an AI agent can consume directly (search, add to basket…) without scraping. Hence a faster, cheaper interaction than raw reading of HTML pages.

These three layers make up the emerging infrastructure of the agentic channel. UCP is the best-established standard to date; ACP and WebMCP are still being adopted.

THE METAPHOR THAT HELPS

Social networks imposed their APIs on brands: if you wanted to exist on Instagram, you spoke Graph API. Agentic commerce imposes its protocols on merchants: if you want to be shoppable by ChatGPT or Gemini, you will speak UCP or ACP. It is not yet a universal condition today, but it is the direction in which the standards are converging. These three protocols answer the same question: how does an AI agent connect to your e-commerce ecosystem — catalogue, stock, orders? That is what matters today, and that is where you should invest.

A further layer is gaining ground: the Agent-to-Agent (A2A) protocol, which lets agents talk to each other, with use cases emerging in IT, supply chain and finance, not yet in e-commerce, even though UCP already provides for A2A as an integration route. Before long, your customer’s Super Agent will be able to address your own Brand Agent directly. A word of caution: this is not the return of yesterday’s shop-window agent. This one does not talk to your customers, it talks to their agent (Gemini, Claude…). It will be a new object in your e-commerce and brand ecosystem.

The confusion comes from a simple fact: several technology layers emerged at the same time, and each carries the same label.

Most of what circulates as “agentic commerce” in 2026 sits between the first two rows. It is not without value, but calling it agentic creates expectations these systems cannot deliver on.

For twenty years, e-commerce has been conceived as a journey, with every step optimised — acquisition, landing page, product page, basket, checkout — to maximise conversion at each point of friction. The agent changes the structure of the problem. In this model the buyer no longer browses: they delegate. They hand over a mandate, and the agent executes. It is no longer the site that persuades; the agent’s ability to interpret, weigh up and act becomes the competitive ground. This agentic channel unfolds in three stages:

• Discovery: the agent explores in the buyer’s place, goes through catalogues, filters on criteria never typed into a form. Your visibility depends on how readable your data is to a model.
• Recommendation: the agent weighs up, compares and ranks the options against the mandate it received. Your credibility depends on how consistent and how fresh your product attributes are.
• Disintermediation: the agent decides, triggers and completes, without the buyer having seen a single merchant interface. Your actionability depends on exposing your transactional capabilities (catalogue, stock, basket, order…) through the standard protocols: UCP, ACP, WebMCP.

This progression is not a forward-looking scenario. It is already under way, and it will have 3 implications that redraw the rules:

1. Classic SEO becomes insufficient

If an agent selects the products before the buyer sees a single page, visibility in traditional engines is no longer enough. Enter GEO (Generative Engine Optimization) and AXO (Agent eXperience Optimization): the ability of your product data to be found, interpreted, judged reliable and acted upon by an AI system. Is your semantic product data readable enough for a model to understand what your product is for and what sets it apart?

2. Loyalty changes its object

A consumer who delegates their everyday shopping stays loyal to the agent that best understands their preferences. Loyalty tends to shift from the brand or the site towards the agentic layer — a hypothesis still taking shape, but a major strategic issue for players who do not control that layer.

3. Price becomes central again, compared continuously

An agent does what a human buyer almost never does: genuinely compare, all the way to the end, instantly. Humans tire quickly, and appearing in the top results was often enough to be chosen. The agent goes to the end of the list. Position no longer protects you; the genuinely best option wins.

*GEO + AXO: from visibility to conversion*

• GEO makes you visible: the AI agent finds you, reads you, consults you. It does not necessarily cite you: you are in the pool of candidates, not yet in the answer.
• AXO makes you *shoppable*: among the candidates it consults, the agent keeps and recommends those it can rely on — stock and prices verifiable in real time, standardised and comparable product attributes, orders executable through the protocols (UCP, ACP)…

Three realistic 6-to-12-month projects that create immediate value and position you for the agentic channel.

1. Audit the agentic readability of your catalogues

Are your product pages structured to be understood by a model?

• Discoverability: can your products be found by an agent that reasons by intent rather than by keyword?
• Readability: are your attributes and descriptions unambiguous, free of internal jargon?
• Credibility: is your data consistent, up to date, free of the contradictions that would make a model doubt?
• Actionability: do your feeds comply with the UCP protocol (the most advanced standard) and anticipate WebMCP?

2. Make your real-time operational data reliable

The catalogue says what the product is; operational data says whether it is available, at what price, within what lead time… An impeccable catalogue backed by a stock feed that is two days old loses you the sale and gets you ruled out by the agent (which remembers). Fresh data, not just clean data.

3. Start on protocol compliance

Assess your exposure to the agentic channel protocols: UCP first, to make your purchase journey actionable by agents; then WebMCP, to expose your tools directly to agents (search, cart…). UCP is already in production at the large platforms: it is the best entry point today. Anticipating now means avoiding the cost of catching up in 12 to 18 months.

Agentic commerce is not one more fad. It is a structural transformation of the relationship between the buyer, information and the act of buying. The dominant channel will not be one agent per brand, but global platforms that hold that channel and brands that make themselves readable within it, or that do not exist there at all. Brand agents will keep their relevance in specific areas, but they will not replace the agentic channel. The short-term challenge will be to make your catalogue, your stock and your transactional capabilities exposed, structured and reliable enough for an agent to draw on them with full confidence. Falling behind will have a cost: lost revenue, falling conversion rates, and churn among the customers an AI agent has steered towards other brands and other products. The brands that adopt the standards of agentic commerce today will also be ready for the next phase: Agent-to-Agent, where their brand agents will talk directly to the customer’s AI assistant to pin down a need, prepare a basket, manage recurring purchases…