Agentic commerce explained, how AI shopping agents are changing ecommerce and website design

Home / Blog / Agentic commerce explained

Picture a customer in Dubai asking an AI assistant to find a premium leather handbag under a set budget, compare the materials, check which sellers deliver to Jumeirah this week, and buy the best match. The customer never opens a browser tab. The agent does the looking, the comparing and, where it is allowed to, the buying.

That is agentic commerce. It is not a prediction any more. Google published an open standard for it in January 2026, Shopify opened its catalogue to AI channels the same day, ChatGPT has had in-chat checkout since late 2025, and Meta launched a personal AI agent this month. Retailers are now making decisions about it in both directions, with Amazon blocking Meta's agent from its site while Shopify and PayPal opened checkout to it, all within one week of September 2026.

For an ecommerce business, this raises a question that sits with the development team more than the marketing team. If a shopper's first contact with your brand is an AI agent reading your product data rather than a person reading your homepage, can that agent understand your catalogue, trust your stock levels and complete a transaction? The answer is decided by your website architecture. This guide explains what agentic commerce is, how AI shopping agents work, why most ecommerce websites are not ready, and what to build so yours is.

What is agentic commerce and what are AI shopping agents

Agentic commerce is online shopping where an AI agent carries out parts of the buying process on a customer's behalf, rather than only advising them. The customer states a goal. The agent interprets it, searches, compares, shortlists, and in supported cases completes the purchase within limits the customer has set.

An AI shopping agent is the software doing that work. The difference from a chatbot is not intelligence, it is permission and action. A chatbot answers, "here are five handbags you might like". An agent checks stock, applies the budget, picks one, and takes it to checkout.

Three levels of AI involvement in shopping

Where the big platforms stand

• OpenAI and Stripe launched Instant Checkout in ChatGPT on 29 September 2025, with the Agentic Commerce Protocol published as an open standard. Merchants stay the merchant of record and keep fulfilment, returns and support.
• Google introduced the Universal Commerce Protocol on 11 January 2026, described as "an open-source standard designed to power the next generation of agentic commerce", built with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by more than 20 partners including Adyen, American Express, Mastercard, Stripe, Visa and The Home Depot. Google had already added agentic checkout features to AI shopping experiences through 2025.
• Shopify announced agentic commerce at scale on the same day, adding native shopping in Google's AI Mode and the Gemini app, an updated Microsoft Copilot integration with embedded checkout, Agentic Storefronts in Shopify Admin, and, for the first time, catalogue access for non-Shopify merchants.
• Meta launched Muse, its personal AI agent, on 8 September 2026. On 21 September, Shopify enabled agentic checkout with Shop Pay for Muse and Amazon blocked Muse from its retail site, stating that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use". PayPal announced Muse checkout support on 22 September.

Two things follow from that list. Agentic commerce has moved from demos to published standards backed by payment networks and large retailers. And it is contested, because every retailer has to decide whether agents are welcome customers or unauthorised traffic.

How AI shopping agents work, from product discovery to checkout

The process is easier to design for when you can picture the six steps.

1. Understanding intent. The agent reads the request and extracts the constraints: budget, category, size, colour, brand preferences, delivery location and deadline.
2. Discovering products. It searches whatever it can reach: product feeds, merchant catalogues, commerce APIs, partner integrations or the open web.
3. Comparing options. It weighs price, specifications, availability, delivery time, return terms and how well each item matches the stated constraints.
4. Selecting. It narrows the list, or presents a shortlist when the customer wants the final say.
5. Transacting. Where an integration, a payment method and the customer's permissions allow, it starts or completes checkout.
6. Post-purchase. It returns the confirmation and, where supported, helps with order status, delivery updates or returns.

How far an agent gets depends on the agent, the merchant's integrations, the payment setup and the approval rules in place. Some agents stop at step four by design. Others complete step five only on merchants that support a commerce protocol. Very few complete all six reliably today.

How agentic commerce changes the ecommerce customer path

The business implications follow from that middle row. If an agent shortlists on structured data, then a product with thin descriptions, missing attributes and unclear delivery terms can be filtered out before a human ever sees your branding. Merchandising still matters, but it stops being the only thing that decides visibility.

There is also a relationship question. When a platform sits between you and the buyer, you risk losing the direct contact that email lists, loyalty and repeat purchase are built on. The counterweight is to keep your own channels strong, which is the argument for a good website and app alongside any agent traffic, not instead of it.

None of this means human shopping is going away. Most UAE ecommerce revenue today comes from people tapping through mobile sites and apps. Agentic commerce is an additional way in, and it should be planned as an extra channel rather than a replacement.

Why most ecommerce websites are not ready for AI shopping agents

A site can be beautiful, fast and easy for people to use, and still be unusable for an agent. The usual reasons are structural.

• Product information is scattered. Key details sit in images, PDFs, tabs loaded by scripts or free-text descriptions rather than in fields.
• Attributes are missing or inconsistent. The same material is written three ways across a catalogue, sizes follow no standard, and variants are half-described.
• Price and stock are hard to read reliably. Values render late, differ between the listing and the product page, or change at checkout.
• Checkout depends on browser behaviour. Multi-step flows, pop-ups, cookie walls and bot challenges break automated interaction.
• There are no documented APIs. Nothing exposes search, cart, checkout or order status in a supported way.
• Variants, shipping rules and availability disagree. A colour shows as in stock nationally but is unavailable for the customer's emirate.
• Bot protection blocks everything equally. Security rules treat a customer's authorised agent the same as a scraper, with no way to allow one and block the other.

The role of website architecture in agentic commerce

Website architecture decides how product information is organised, how systems talk to each other, and how a transaction is validated before money moves. Those three decisions determine whether an agent can understand, trust and transact with your store. Our ecommerce website development work starts in the same place for the same reason.

Structured product data and catalogue architecture

Treat the catalogue as a database, not as a set of pages. Every product needs a stable identifier, a consistent name, a clear description, complete attributes, defined variants, accurate price and availability, and delivery information. Attributes should use consistent values, so "genuine leather" is not competing with "real leather" and "leather (full grain)" in the same catalogue.

Publish that data in machine-readable form as well as on the page. Product schema markup, clean product feeds and consistent values across the site, the feed and the API are what let an agent match your product to a customer's constraint. If the page says one price and the feed says another, an agent has no way to decide which to believe, and a mismatch is a reason to drop you from a comparison.

API-first and integration-ready architecture

An API-first build exposes commerce capability as services: product search, product detail, inventory, cart, checkout and order status. That is what lets a website, an app, a marketplace integration and an authorised agent all use the same logic and the same data, instead of each one reimplementing the rules.

Design the integrations around standards that the platforms you care about actually support, and verify support rather than assuming it. An agent cannot connect to an API just because the API exists. It needs a supported route, credentials and permission.

Headless and composable ecommerce architecture

Separating the presentation layer from commerce services gives you flexibility to serve several interfaces at once, including a website, a mobile app and authorised AI channels. That is the practical case for headless or composable builds in this context.

It is not mandatory. A well-built platform store with clean data, solid feeds and a supported commerce integration can serve agents perfectly well, and it is usually the faster path for a small or mid-sized retailer. Headless earns its place when you have several front ends, complex catalogue logic or integration requirements that a single-platform template cannot carry.

Real-time inventory, pricing and availability

An agent making a recommendation on a stale price creates a bad outcome for everyone: the customer is misled, the merchant deals with the complaint, and the platform loses confidence in the merchant. Stock should be accurate at variant level, prices should reflect the customer's market and currency, and the values used at checkout should be validated at that moment rather than read from a cached feed.

This is where ERP integration earns its keep. If stock lives in an ERP or POS and reaches the website once a night, the website will disagree with reality several times a day. Agents surface that problem faster than shoppers do, because they check.

Secure checkout, payments and transaction validation

Agentic checkout should never mean handing an agent raw card details. The published protocols are built around tokenised payments, delegated authority and explicit customer consent for a specific transaction. On your side, that means a payment gateway integration that supports tokenised and delegated flows, server-side validation of price, stock and shipping before authorisation, fraud controls that can tell an authorised agent transaction from a suspicious one, and a clear confirmation record.

Spending limits, approval steps for higher-value orders and a visible audit trail belong in the design, not in a later phase.

Scalability, monitoring and error handling

Agent traffic behaves differently from human traffic. It can arrive in bursts, retry aggressively and hit the same endpoints repeatedly. Plan for rate limits, idempotent order creation so a retry does not create two orders, graceful responses when a product is unavailable, clear error codes instead of generic failures, and logging that lets you see what an agent requested and what it received. Without those logs, diagnosing a failed agent order is guesswork.

How to design an AI-ready ecommerce website

The architecture points above turn into a practical checklist.

• A structured, complete product catalogue with every commercially relevant attribute filled in.
• Stable product identifiers and variant data, consistent across the site, feed, API and back office.
• Machine-readable product information, including product schema markup and a clean, current product feed.
• Accurate, accessible pricing and inventory, at variant level and by market.
• Documented, secure APIs or a supported commerce integration for search, cart, checkout and order status.
• Product discovery that works without a human, with filters and search logic that map to real attributes.
• Secure cart and checkout workflows, including tokenised payments and server-side validation.
• Clear authentication, consent and authorisation controls, so you can permit an authorised agent without opening the door to everything.
• Reliable order confirmation and status, available through the same interfaces.
• Monitoring, logging and performance work, so agent traffic is visible and measurable.

One caution. Schema markup and crawlability help discovery, and they are worth doing regardless. They do not make a store transaction-ready on their own. Whether an agent can actually buy from you depends on the platform, the integration, the permissions and the commerce standard involved.

AI agent compatibility, design, UX, SEO and technical readiness

Why user experience still matters

Humans remain the majority of your buyers, and they are also the ones who approve what an agent proposes. Clear navigation, honest product information, mobile responsiveness, transparent delivery and returns policies, and a short checkout all still decide conversion. Good website design is not in tension with agent readiness. Most of what makes a page clear to a person also makes it clear to a machine.

Technical SEO and machine readability

Crawlability, semantic HTML, clean product URLs, correct canonical tags, structured data and consistent information across pages are the foundation for both search engines and AI systems. This is ordinary technical SEO and website optimisation work, extended to a new consumer of the same data.

Keep one distinction in mind. Search crawling is anonymous, read-only and public. Agent commerce is often authenticated and transactional. Being easy to crawl helps an agent find and describe you. It does not authorise anyone to buy. Our ecommerce SEO work covers the discovery half of this, including how products are described so AI systems can quote them accurately.

Website performance and reliability

Page speed, stable endpoints, uptime and reliable API responses support everything above. An endpoint that times out under load will fail an agent in the middle of a transaction, and the agent will not retry politely forever. Performance alone does not make you agent compatible, but poor performance will undo the rest of the work.

Agentic commerce standards and integrations to know

These solve different problems and are not interchangeable. A protocol describes how systems talk. Your architecture decides whether you have anything worth saying: clean data, accurate stock, a checkout that validates. Check what your platform and payment provider support today before committing to an integration, because support is moving quickly and announcements run ahead of availability in many markets, including this one.

Risks and challenges of AI-powered shopping

This is not a one-sided opportunity, and a balanced view is the responsible way to plan.

• Fraud and payment security. On 22 September 2026, banks including NatWest, Bank of America, ING, Capital One and Commonwealth Bank of Australia warned that AI shopping agents raise scam, fraud and data-privacy risks, pointing to agents entering card details into websites and steering customers toward payment methods with weaker protections. They plan to take proposals to policymakers, including disclosure when a transaction is made by an agent.
• Unauthorised access. Merchants have to decide which agents may transact and enforce it, as Amazon did when it blocked Meta's agent.
• Wrong recommendations from stale data. An agent working from an outdated feed will promise what you cannot deliver.
• Consent and data privacy. Customer data passing through an agent raises questions you should answer in your privacy policy and check against UAE data protection requirements.
• Loss of the direct relationship. If a platform owns the conversation, your brand can become a line in a comparison table.
• Integration complexity and maintenance. Protocols and platform requirements are changing every few months.
• Unclear attribution. Analytics were not designed for agent-driven sessions, so measuring the contribution of this channel is difficult today. One data point on demand: British retailer John Lewis reported that AI-originated searches rose from 0.3% to 2.5% year on year.

How to prepare your ecommerce website for agentic commerce

1. Audit what you have. Review the current architecture, catalogue quality, feeds, APIs, stock accuracy and checkout. Most stores find data problems before they find integration problems.
2. Fix the product data. Complete attributes, consistent values, correct variants, clean descriptions, accurate delivery information and valid schema markup.
3. Review integrations and protocols. Establish what your platform, payment provider and marketplaces support now, and what is on their roadmap.
4. Strengthen stock, pricing and validation. Synchronise inventory with the back office, validate at checkout, and add tokenised payment support.
5. Test in a controlled way. Try supported agent discovery and purchase flows in a staging environment, and decide your policy on agent access before traffic arrives.
6. Measure. Monitor errors, API performance, product visibility in AI answers and any AI-referred traffic, and review it monthly.

Nobody needs to rebuild everything this quarter. Start with an architecture and data assessment, then sequence the work by what your platform supports and what your customers actually buy.

Building for human shoppers and AI agents

Agentic commerce is likely to change where the first impression happens. For some purchases it will move from your homepage to a data comparison you never see, made by software acting for your customer. The businesses that do well in that setting will not be the ones with the loudest banners. They will be the ones whose product data is accurate, whose stock is real, whose checkout validates properly and whose integrations are supported.

That work pays off either way. A catalogue with complete attributes converts better for people too. Accurate stock reduces refunds. A validated checkout reduces failed orders. A clean API lowers the cost of your next app or marketplace integration. This is the rare case where preparing for an uncertain future means fixing things that are already costing you money today.

If you are planning an ecommerce build or replatform in the UAE, it is worth designing the architecture with both audiences in mind now, rather than retrofitting it in two years. Tomsher builds ecommerce websites, apps and integrations from our Dubai office with a fully in-house team, and our web application development work covers the API and integration layer that agentic commerce depends on.

Frequently asked questions

Sources

• Google Developers Blog, Under the hood, Universal Commerce Protocol
• Shopify, AI commerce at scale
• OpenAI, Instant Checkout and the Agentic Commerce Protocol
• Meta Newsroom, Introducing Muse
• TechCrunch, Amazon blocks Meta's AI agent
• Reuters via CNBC, Banks warn AI shopping bots raise scam, fraud and data-privacy risks