What Is Agentic Commerce? The Complete Guide

Agentic commerce is commerce in which AI agents act, not only advise. A shopper hands a goal to an agent, and the agent searches, compares and sometimes pays. A merchant hands growth work to agents, and they find the problem, build the response and wait for a person to approve it. Both sides are moving at the same time, and most of what I read on the subject covers only one of them. Last updated: September 2026.

I have spent 13 years in eCommerce. Most shifts I watched in that time changed one side of the counter. Search changed how people find a store. Mobile changed where they buy. Agentic commerce changes who does the work on both sides of the same transaction, and that is why it deserves a full map instead of a hot take.

I also have a stake in the merchant side, so here is the disclosure up front. Nexus by Omniconvert is the AI eCommerce growth engine for Shopify brands: it reads your store, ad and customer data, ranks growth opportunities by profit impact, and builds the campaigns to act on them, and you approve what goes live. Later in this guide I am specific about where it fits and where it does not.

If you want the short, question-by-question version focused on data, read agentic commerce readiness: 16 questions answered. This page is the complete guide: the definition, the two sides, what it is not, what changes in product data, checkout and trust, and a seven-step framework you can start this quarter.

What is agentic commerce?

The word that carries the meaning is agentic. An agent is software that pursues a goal over several steps, chooses its next step itself, and uses tools to act. It can search a catalogue, open a product page, fill a cart or call a payment service. A model that only answers a question is not an agent, however good the answer is.

Three properties separate an agent from the AI features most stores already run:

• It has a goal, not a prompt. Find trail running shoes under a set budget that arrive before Saturday is a goal. Which running shoes do you sell is a prompt.
• It takes actions. It clicks, calls an API, adds to cart, pays or publishes. The output is a change in the world, not a paragraph of text.
• It works inside delegated limits. Someone sets the budget, the acceptable brands and the approval rules. An agent is only as safe as those limits are clear.

Put those together and you get a definition you can test against any vendor claim: agentic commerce is commerce in which software acts for a buyer or a seller, inside limits a person set, and a person stays accountable for the outcome.

The term gets used loosely, so one more note. Some people mean only the shopper side, AI assistants that buy things. Others mean the merchant side, AI that runs marketing work. Both uses are correct. The useful view holds both, because the two sides change the same store at the same time.

The two sides of agentic commerce

Consumer side: shopping agents

A shopping agent acts for one person. The person states an intent, a budget and some preferences. The agent searches across stores, reads product data, compares total cost and delivery, and returns a short list. Where the person allows it, the agent also completes the purchase.

Shopping agents reach your store in two ways. Some drive a website the way a person does, through a browser: they load pages, pick a size and fill in the checkout form. Others skip the page and read structured sources, such as your product feed, your schema markup, or a checkout interface that your commerce platform or payment provider exposes. The first kind trips over the same friction your customers do. The second kind never sees your design.

Either way, your counterpart in that moment is software. It has no brand memory and no patience for ambiguity. It also has no reason to accept a risk when three other stores state their terms clearly.

Merchant side: growth agents

A merchant-side agent acts for the store. It reads store, ad and customer data, spots a problem or an opportunity, drafts the response and hands it to a person to approve. The response can be a test, a customer segment, a set of ad creative, a landing page or an email.

This side matters because of what it replaces. In most teams I speak with, skilled people act as the glue between tools: export here, clean there, upload somewhere else. I call that the Human Middleware Problem, and by Omniconvert's estimate eCommerce managers spend about 3 hours a day assembling data. Merchant-side agents take over the assembly. The team stops producing every output by hand and starts deciding which outputs run.

The two sides meet in one place: your data. The shopping agent reads what your store publishes. The merchant agent reads what your store records. If either is incomplete, the agent on that side works from a wrong picture of your business.

What agentic commerce is not

Vendors now attach the word agentic to almost everything, so it helps to know what does not qualify. None of the tools below are bad. They belong to a different category, and buying one as if it were an agent sets the wrong expectation.

• A chatbot. A support or shopping chatbot waits for a question and returns an answer inside one conversation. It may run on a strong language model. It still does not go and do the task. When the shopper closes the tab, the work stops.
• A recommendation widget. A customers-also-bought carousel ranks products for a person who still decides, clicks and pays. It works on one page of one store, and nobody delegated a goal to it.
• A rules-based flow. An abandoned-cart email fires on a trigger and follows a fixed path. That is useful automation, but it does not choose its own steps toward a goal.
• AI search visibility alone. Being named when an assistant answers a question is a prerequisite. Agentic commerce starts when the assistant acts on the answer. The preparation overlaps, but it is a different job.

A quick test works in any product demo. Ask three questions. Does it act, or only answer? Does it pursue a goal across more than one step? Does it work inside limits that someone delegated and can audit? Three yes answers describe an agent. Anything less is a feature with a new label.

Product data that agents can read

A person on your product page does a lot of quiet repair work. They zoom into a photo to check the fabric. They scroll to find the size chart. They see a free-shipping banner and assume it applies to them. An agent does none of that repair work. It reads fields, and it treats a missing field as a missing fact.

That changes what good product data means. Four things matter most:

• Complete attributes. Brand, identifiers such as GTIN, materials, dimensions, compatibility and care. A shopper asks an agent for a waterproof jacket that packs small. If the waterproof rating and the packed size are not fields, your jacket is not in the comparison.
• Clean variants. Each size and colour carries its own price, stock status and identifier, linked to one parent product. An agent buys a specific variant, not a product family.
• Terms as fields. Shipping cost, delivery window, return window in days and who pays for return shipping. Schema.org defines types for this, including Offer, OfferShippingDetails and MerchantReturnPolicy, and your markup should use them.
• Agreement everywhere. The price and stock on the page, in the feed and in the markup must match at the moment the agent checks. A mismatch reads as risk, and an agent that compares several stores can simply drop the risky one.

Then there is the copy. Agents match intent, so descriptions that answer real questions serve them better than descriptions that only set a mood. Keep the brand voice for people, and put the facts in plain sentences next to it. In my experience the two goals conflict far less often than teams fear.

Checkout and payments that agents can complete

A browser-driven agent meets your checkout exactly as a customer does. Every pop-up, forced login and last-step fee is an obstacle. Baymard Institute puts average documented cart abandonment at about 70%, and the reasons shoppers give are familiar: extra costs revealed late, a required account, a process that is too long. An agent hits the same walls, and it has less reason than a person to push through them.

A protocol-driven agent skips the form. The merchant side exposes checkout through a defined interface, and the agent sends a structured order with a payment credential. The standards for this are forming in public. In 2025, OpenAI and Stripe published the Agentic Commerce Protocol, and Google published the Agent Payments Protocol (AP2), which focuses on proving that a person authorised the payment an agent makes. Card networks and payment processors are building their own programs for agent payments too.

My advice here is conservative. Do not build a custom integration for each protocol while the standards settle. For most Shopify brands, the practical path is to adopt what your commerce platform and payment provider support, and to spend your own effort on what every route needs:

• Guest checkout that works without an account.
• Total cost, including shipping and taxes, visible before the last step.
• Discounts that exist as real codes or automatic rules, not only as banner images.
• Bot protection that separates fraud and scraping from legitimate agents, instead of blocking every automated visitor.
• Order confirmation, tracking and returns that work without a phone call.

Trust, approval and control

Every agent works on borrowed authority. The question on both sides of the counter is the same: how much authority, and who checks the result?

On the shopper side

A person who delegates a purchase sets a mandate: a budget, acceptable brands, a delivery date, and often a rule to ask before paying. Agent payment programs aim to carry those limits with the credential, so the agent can pay without holding a raw card number. For you as a merchant, the effect is practical. The clearer your terms, the easier the shopper's approval is, and an order that matches what the shopper authorised is an order that stays sold.

On the merchant side

Your store now has three kinds of automated visitors: agents buying for real people, crawlers reading for search or model training, and bots scraping prices or attempting fraud. A blanket block on automation stops all three. Review your bot rules with those groups in mind, and make the decision about training crawlers separately from the decision about agents that shop. Disputes also get more precise. When an agent picks the wrong size, the case turns on what your data said, so clear terms protect you.

Inside your own team

The same logic applies to merchant-side agents. Write down which actions an agent may take alone, which need a person's approval, and how the review loop works. A team that approves everything by hand becomes the bottleneck again. A team that approves nothing loses control of its brand. The executor-to-supervisor operating model sets this out as four control surfaces: the objective, the constraints, the approval threshold and the review loop.

An agentic commerce readiness framework in seven steps

Here is the full sequence I would run, in order. Steps 1 to 5 prepare your store for shopping agents. Steps 6 and 7 prepare your team for merchant-side agents. Every step pays off even if agent checkout grows slowly, because people and search engines reward the same work.

1. Audit what assistants say about you. Take the ten questions customers ask you most and put them to the assistants they use. Record every wrong price, wrong stock status and outdated policy. Log each one as a bug with an owner.
2. Create one source of truth for price, stock and policy. Decide which system owns each fact. Make the page, the feed and the structured data read from that system, and check how often each one refreshes.
3. Turn implied facts into fields. Start with your best-selling products. Every fact a customer gets from a photo, a banner or a review should exist as an attribute. Fix variants so each one carries its own price, stock status and identifier.
4. State purchase terms in numbers. Publish shipping cost, delivery window, return window and return shipping rules in days and currency. Add the matching markup and confirm it matches the visible page.
5. Remove checkout friction. Allow guest checkout, show total cost early, make every discount a real rule, and tune bot protection to stop fraud rather than every automated buyer. Adopt agent checkout through your platform and payment provider, not through custom builds.
6. Measure customers, not sessions. An order an agent places can arrive with little or no journey data. Build measurement on what does not disappear: customer value, RFM segments and profit after costs. True Profit is a better guide here than ROAS.
7. Put your growth work under supervision. List every place a person moves data between tools by hand. Give the most time-consuming one to a merchant-side agent, and write the approval rules before the agent ships anything.

For the data steps, the readiness FAQ answers the detailed questions, such as whether you need a separate feed and whether to block AI crawlers. A conversion audit makes a good companion to steps 3 and 4. It asks whether a person can find and trust the facts on a page, and agent readiness asks whether software can read the same facts. Run both checks on the same pages, because one fix often serves both readers.

Where Nexus by Omniconvert fits

Nexus is an AI eCommerce growth engine for Shopify brands. It runs 840 AI agents on the merchant side of the growth loop, and the work splits into five parts:

• It reads. It unifies store, ad and customer data and builds RFM segments, so decisions start from customer value rather than session counts.
• It ranks. It ranks growth opportunities by profit impact, so the first item you see is the one most likely to move margin.
• It builds. It generates launch-ready assets from customer data: static and video ad creative, landing pages and email copy. It also monitors competitor ad libraries and classifies competitor creative by angle, hook and offer.
• It measures. It reports results in True Profit, not ROAS.
• You approve. You approve what goes live, so every campaign passes a person before it reaches a customer.

Two boundaries matter for this guide. First, Nexus is not a shopping agent, so it does not buy on behalf of your customers. Second, the storefront work in steps 2 to 5 (feeds, markup and checkout) stays with your commerce platform and your feed tools. Nexus covers steps 6 and 7: customer-level measurement and growth work you supervise. It leaves the bid and campaign lifecycle to the ad networks.

On the Shopify App Store, Nexus holds 5.0 out of 5 from 60 reviews (as of September 2026). For the full category definition and a buying scorecard, read what an AI eCommerce growth engine is. If your slowest step is getting from a customer insight to finished creative, the creative brief bottleneck guide covers that step in depth.

FAQ: agentic commerce

What is agentic commerce in simple terms?

Agentic commerce is buying and selling in which AI agents act, not only advise. A shopper can ask an agent to find, compare and buy a product within a budget. A merchant can use agents to find growth opportunities and build campaigns. In both cases a person sets the limits and approves the decisions that matter.

How is agentic commerce different from conversational commerce?

Conversational commerce is shopping through a conversation, such as a chatbot or a messaging app, where a person still makes each decision. Agentic commerce hands part of the task to software that acts. The agent searches, compares and can complete a purchase inside limits the person set, without a new message for every step.

Do AI agents already buy from online stores?

Research and comparison through AI assistants are already common. Agent-completed checkout exists, but it is early. In 2025, OpenAI and Stripe published the Agentic Commerce Protocol and Google published the Agent Payments Protocol, and the standards are still settling. Prepare your data and checkout now, because the same work helps search visibility and human conversion.

Should I build an integration for every agent checkout protocol?

Not yet. The standards are still settling, and a custom build for one protocol is an expensive bet on one version of the future. Use the agent checkout options your commerce platform and payment provider support. Spend your own effort on accurate product data, clear purchase terms and a checkout without needless friction.

What is the difference between a shopping agent and a merchant-side agent?

A shopping agent acts for a person: it searches stores, compares offers and buys within a budget. A merchant-side agent acts for the store: it reads store, ad and customer data, ranks opportunities, builds campaigns and measures results. The first changes who reads your store. The second changes who does your growth work.

How does Nexus by Omniconvert fit into agentic commerce?

Nexus by Omniconvert works on the merchant side. It reads your store, ad and customer data, ranks growth opportunities by profit impact, and builds the campaigns to act on them, and you approve what goes live. It is not a shopping agent, and your feed, markup and checkout readiness stay with your commerce platform.