Artificial Intelligence

Agentic commerce: what e-commerce SMEs must prepare in 2026

Arnaud Guedj·October 12, 2026·7 min

Cheerful retro orange robot at a miniature checkout while small clay robots pass parcels and cart icons, a metaphor for agentic commerce.

Agentic commerce starts when AI stops simply answering a question about a product. It gets a mission, say “find me a two-person tent under 200 euros, delivered before Saturday”, and works through the steps on its own: searching, comparing, checking stock, sometimes placing the order. For an e-commerce SME in 2026, the point is not to bolt another bot onto your site. It is to make your catalog, inventory and rules readable by these agents before traffic goes elsewhere.

Chatbot or agent: not the same thing

People mix the two up all the time, and that is where many projects go wrong.

A chatbot answers. You ask a question, it gives you an answer, then waits for the next one. “Is this jacket waterproof?” It reads the product page and says yes or no. That is all. You decide what happens next, and you do the clicking.

An agent acts. It receives a goal, breaks the work into tasks and carries them out one after another using tools: a site’s search engine, a stock API, a price comparison, a cart, sometimes a payment method. It checks the result of each step and adjusts. If size M is out of stock, it looks at another model or another retailer without being asked again. It comes back to the customer with a proposal, or with an order placed within the limits it was given.

Take the same need on both sides. With a chatbot, the shopper asks five questions about five products and does the comparison in their head. With an agent, they state the need once, and the agent fetches the five products, compares them, drops the ones that cannot arrive in time and presents the best two.

For a merchant, the difference is very concrete. A chatbot reads your site like a hurried human. An agent queries your data: it needs precise attributes, accurate stock and clear rules (lead times, returns, delivered price) to chain its tasks without mistakes. A chatbot on top of a messy catalog gives vague answers. An agent facing a messy catalog gives up, or picks a competitor.

What agentic commerce changes in the buying journey

Part of discovery, comparison and sometimes the transaction itself is handed to an AI agent: a feature inside ChatGPT or Gemini, a marketplace assistant, or a merchant’s own agent. The shopper describes a need in plain language, the agent does the searching and sorting, and in some cases triggers payment.

At Decathlon, Ismael Ghozael, VP Product & Platforms, described three phases in a discussion reported by EcommerceMag on October 5, 2026. First, attention moves to AI interfaces. Then comes commerce itself: who actually completes the purchase? Finally, several agents coexist, the consumer’s and the merchants’, dealing with each other. His conclusion is down to earth: “Ultimately, it’s an additional sales channel.” One more channel, to be prepared like the others.

Why act in 2026

Product discovery has already partly moved to AI, even though agent-led payment is still rare. According to Minted (January 7, 2026), studies by Toluna, OpinionWay, Iligo and BCG show that between a third and half of internet users already use generative AI when shopping. The Iligo / Media Figaro barometer finds that 70% of French respondents have followed an LLM recommendation, and 42% have bought a suggested product. OpinionWay (October 2025) counts 29% who use AI to research a purchase, 47% including those who plan to. BCG, covering Black Friday 2025, reports that 48% of 10,000 consumers surveyed in ten countries used or planned to use GenAI for year-end shopping, up from 39% in 2024.

Be careful about what these numbers measure: mostly chatbot-style use, where people ask questions to get informed. Agents that actually buy on the shopper’s behalf are still a minority.

On traffic, Adobe (cited by Minted) found that visitors coming from AI platforms converted 38% better than others. Digiday, on the other hand, notes that ChatGPT still accounts for only 0.2% of total traffic to major U.S. e-commerce sites, 200 times less than Google. The volume is small, but the people arriving this way know what they want.

At the One-to-One Retail E-Commerce event in Monaco (March 2026), Sophie Marchessou (Mirakl) told FashionNetwork that some brands already get 10 to 15% of their traffic from LLM searches, while sales completed directly on those platforms remain “very anecdotal.” Marc Lolivier (Fevad) pointed to the psychological brake: shoppers want to keep control of payment, much like the early days of e-commerce when people hesitated to type in a card number.

Making your catalog usable by an agent

For an SME, the most concrete lever is not adding a chatbot to the homepage. It is your product feed, because that is where an agent finds the material it needs to chain its tasks. Grégoire Chantry (Arcane Groupe / Cosmo5), in a Journal du Net column dated July 16, 2026, calls the product feed the underrated asset of agentic commerce: it drives visibility in LLMs, AI Ads and “zero-click” buying. Queries are getting longer, and models use every attribute available: description, highlights, material, size, availability.

Two layers add up:

The descriptive layer: titles, attributes, stock, price and selling points, structured and kept up to date. This is what lets an agent filter and compare without errors.
The conversational layer: shoppers’ frequent questions and objections, with answers built into the feed, ready to be reused. The column mentions OpenAI’s Agentic Commerce Protocol here.

Anne-Claire Baschet (Mirakl), quoted by Minted, sums up the problem: users speak in natural language, with context and a goal. In August 2024, only 4% of LLM users wrote queries longer than eight words. By April 2025, 41% did. Your product pages need to say who the product is for and what problem it solves, not just show a reference number.

B2B shows the gap between answering and acting quite clearly. Visionary Marketing (September 21, 2026, updated September 30) describes DJUST agents that take an email or an Excel file, extract the lines, check the references and prepare a structured order for an operator to approve. No chatbot does that: it means reading, checking, keying in and passing on, several tasks in a row. Arnaud Rihiant notes that around 80% of B2B sales teams still lose time re-entering orders by hand into the ERP. I see the same thing in many SMEs: the blocker is not AI, it is the data.

Payment and trust

As long as AI only answers, a mistake costs a bad piece of information. When an agent chains tasks all the way to payment, a mistake can cost an order. Payment becomes the checkpoint. In the Decathlon / Purse discussion (EcommerceMag, October 2026), Pascal Roche stresses the chain of players (merchant, payment provider, bank, authentication, fraud prevention) and the risk of blurred responsibility. Ismael Ghozael is firm: an algorithmic intermediary must not dilute the merchant’s responsibility. Trust comes down to execution: accurate stock, delivery on time, easy returns.

For an SME, that gives three rules:

Do not open agent payment if your displayed stock or lead times are approximate.
Keep a record of who authorized the transaction: the customer, the agent or a business rule.
Let a human decide ambiguous cases, as DJUST recommends in B2B.

My prescription for an e-commerce SME

Don’t confuse the projects. A chatbot on your site improves customer service. Being ready for agents is something else: making your data usable by an AI that chains tasks on the shopper’s behalf. You can do both, but not with the same budget or priorities.
Treat your product feed as a sales channel, not a technical export: complete attributes, real-time stock, answers to common questions, starting with the 20% of products that make 80% of revenue.
Measure traffic from AI now (ChatGPT, Gemini, Perplexity, marketplace assistants): volume, landing pages, conversion, even if it is still small.
Work on your reputation beyond your own site (reviews, press, forums, Wikipedia): LLMs rely on outside sources, not only your product pages.
Separate discovery from payment. Be visible to agents now, and automate payment only with an audit trail and clear responsibilities.
Keep the customer relationship. After-sales, delivery and loyalty are still your ground. Fabien Versavau, quoted by Minted, puts it bluntly: agents hallucinate, and your customers stay loyal if you treat them well.

A chatbot helps you answer your customers. An agent goes shopping for them, and it will pick the merchants whose data lets it finish the job. SMEs that prepare in 2026 won’t win the agent payment battle yet. They make sure they are found, understood and chosen on the day the shopper hands over the search.

Frequently asked questions

What is the difference between a chatbot and an AI agent?

A chatbot answers a question and waits for the next one: the shopper decides and clicks. An agent receives a goal and carries out the tasks itself (searching, comparing, checking stock, filling the cart, sometimes paying) using tools, until it comes back with a proposal or an order.

Will agentic commerce replace my online store?

No. Practitioners describe it as an additional channel. In March 2026, Mirakl noted that sales completed directly on LLMs remain anecdotal, even if some brands already get 10 to 15% of their traffic from AI searches. Your site still owns trust, after-sales and loyalty.

Where should I start with a limited budget?

With your product feed: complete attributes, reliable stock and answers to shoppers’ questions. That is what an agent needs to chain its tasks. Then measure traffic coming from AI assistants. There is little point opening agent payment if machines cannot yet understand your catalog.

Should I already accept payment through an AI agent?

Not necessarily. Fevad points out that shoppers still want to control payment. Be visible to agents for discovery first, and automate payment only with a clear audit trail and well-defined merchant responsibility.

What is GEO in this context?

GEO (generative engine optimization) means shaping your content and data so you show up in answers from generative engines and agents. Unlike classic SEO, it also depends on outside sources (reviews, press, forums) and on conversational answers built into the product feed.

Does B2B matter as much as B2C?

Yes, from a different angle: agents that turn an email or a spreadsheet into a structured order, with human approval. Visionary Marketing (September 2026) describes this with DJUST: the agent reads, checks and prepares the order, and an operator approves it.