Where agentic commerce goes next: agents, voice and AR

Agentic commerce means AI agents doing part of the work of buying and selling: finding and comparing products for shoppers, and preparing changes to content, prices and stock for a store's team to approve.

Some of it is already live. Shopify and BigCommerce both provide MCP servers that give AI assistants a standard route into a store, Shopify has added checkout tools for browser agents, and Chrome is shipping a browser standard called WebMCP behind a flag. Most of the rest is still a forecast. Coordinated teams of agents, agents that improve from your team's feedback, voice shopping and agent help inside AR are all early.

What to do now is practical. Make your product data accurate and consistent, switch on the agent routes your platform already offers, and decide who approves what before any agent changes your store. The guide ends with a short list to work through.

What has already changed

When we first wrote about agentic commerce in 2025, most of it was prediction. We expected ecommerce platforms to adopt the Model Context Protocol (MCP), the open standard introduced by Anthropic that gives an AI application one way to connect to the tools and data in another system. Some of that is now in place:

Shopify provides Storefront MCP and Catalog MCP for eligible stores.
BigCommerce has offered its own Storefront MCP server to every live store since 11 May 2026.
On 28 September 2026, Shopify added checkout tools for browser agents. They run inside the checkout page, and the merchant does not need to set anything up.
Chrome is shipping WebMCP, a browser standard that lets a web page offer tools to an assistant. As of October 2026 it sits behind a flag in stable releases, so most shoppers do not have it switched on yet.

Two routes are forming. MCP servers connect an assistant to the store from outside. Browser tools, such as WebMCP and Shopify's checkout tools, work inside the page the shopper has open. We expect most stores to need both. MCP for ecommerce covers the first route and WebMCP for ecommerce covers the second. For dates, platform details and a checklist, see the agent-ready store audit.

The basics have not moved. An agent can only recommend or buy what it can read, so accurate titles, prices, stock, delivery and returns information still come first.

Multi-agent systems and how agents coordinate

A multi-agent system splits work between several agents, each with one job, instead of asking one general agent to do everything. In a store, that might be one agent for product content, one for images, one for search visibility, one for stock and one for testing. Each is easier to check because its job and its permissions are narrow.

The hard part is coordination. Agents need a shared picture of the store: what is in stock, what changed yesterday, which campaign is live, and what your team has already approved or turned down. MCP helps because every agent can reach the same tools and live data in the same way, instead of each keeping its own copy. It does not decide who does what. That still needs rules, and people to set them.

An example: launching a new range

Say you are adding a new range that will sell in several countries. Spread across agents, the work might run like this:

An audit agent checks the new product pages for missing titles, descriptions and attributes, and for broken links.
A content agent drafts titles and descriptions in your brand voice.
An image agent converts photos to a lighter format, such as AVIF, and checks they still look right.
A translation agent prepares each language version from the approved copy.
A testing agent sets up a test on the category banner and reports which version did better.

Each step produces a proposed change, and the proposals arrive in one queue for your team to review, instead of five tools each writing straight to the live store. Order matters too: translating a first draft means translating it again once the copy changes.

When agents disagree

Agents with different goals will clash. A stock agent may want to hide a product that is running low while a merchandising agent is promoting it on the home page, or a pricing agent and a promotions agent may both try to change the same price. Before you run more than one agent, agree a few rules:

Give each field or action one owning agent. Only the pricing agent proposes price changes.
Keep one shared log of what every agent proposed, what was approved and what was applied.
Set limits on how many products one change can touch and how far a price can move.
Send conflicts to a person, with both proposals side by side.

Our forecast is that, over the next two years, more ecommerce software will arrive as groups of specialist agents. The teams that get value from them will be the ones that settle ownership and approval rules early.

Agents that learn from feedback

In 2025 the promise, ours included, was agents that keep learning and grow with your business. Some of that is real, but "learning" usually means something more modest than it sounds. In most agent products it means one of these:

The agent is given a new skill: support for a new image format, a connection to another analytics source, a new kind of report.
The agent keeps a record of what your team approved, edited or rejected, and uses it as guidance next time. If your team always shortens product titles, the next drafts should come in shorter.
The people running the agent review its results and tighten its instructions and checks.

An agent retraining its underlying model on your store's data overnight is rare. That is no bad thing. Changes you can see and review are easier to trust than a model that quietly shifts.

Why approval and undo still matter

An agent that learns can learn the wrong lesson. If a discount once lifted sales of a slow line, an agent may keep proposing discounts long after the reason has gone. The more an agent adapts, the more you need to see what it is about to do. A few habits keep learning agents safe:

Approval before anything changes the live store, spends money or reaches a customer. The person who approves should not be the person, or the agent, that prepared the change.
An undo for every change, agreed before it is applied, and where you can, a test on a copy of production first.
A reason attached to every rejection. A one-line reason is the most useful feedback an agent can get, and it leaves a record your team can check later.

Approval is also how an agent learns what good looks like in your store. Take the approval step away and you lose the feedback with it.

Voice commerce

Voice commerce means shopping by speaking to an assistant on a phone, a smart speaker or in a car. It suits some purchases much better than others. Reordering something you already buy, checking where an order is, or asking a quick question about delivery all work well by voice. Choosing a sofa, or comparing three pairs of trainers, mostly does not, because people want to see what they are buying.

What changes with agents is what sits behind the voice. A voice assistant that can use the same tools as a text assistant can search a catalogue, check stock and build a basket, so voice becomes one more way into the same agent. For a store, being ready for voice is mostly about data:

Product names that make sense read aloud, without internal codes or long strings of keywords.
Variants named clearly (size, colour, pack size), so "the large blue one" points to a single product.
Delivery costs, delivery times and returns written plainly, so an assistant can repeat them accurately.

Voice also helps shoppers who find screens hard to use, which is reason enough to keep product information clear and complete. Our forecast is that, over the next two years, voice grows first in repeat purchases and customer service and stays a minor route for considered purchases. Treat it as a test, not a channel to plan revenue around.

AR and visual shopping

Augmented reality (AR) lets a shopper see a product on themselves or in their room through a phone camera. Try-on and view-in-your-room previews are not new. The 2025 idea was to put an agent inside that view: you point your phone at your living room, and an assistant suggests a lamp that suits the space, then adds it to your basket.

Agent-led AR shopping is early. It needs a 3D model, accurate measurements and fast rendering for every product it covers, which is a large job for a big catalogue. It also points a camera at people's homes and faces, so shoppers need to know what is captured and whether anything is stored.

Visual search is the nearer change: a shopper photographs something they like and asks an assistant to find it, or something similar. Your product images and attributes are what get matched. The groundwork costs little and helps every channel:

Dimensions, materials and colours held as product data, not only shown in images or written into the description.
Clear photos from several angles, with descriptive alt text.
If AR suits your category (furniture, eyewear, home decor), a small test on a few best sellers before anything larger.

Our forecast is that AR stays a feature for categories where size and fit decide the sale, and that agent help inside AR remains at the test stage for most stores over the next two years.

How shopper behaviour is changing

Personalised recommendations have shaped online shopping for years. The bigger shift now is that shoppers increasingly meet stores inside AI answers, and some hand over part of the job: find the options, compare them and build the basket.

From searching to delegating

Picture a shopper asking an assistant for a waterproof jacket under a set budget that will arrive by Friday. The assistant does the comparison that used to happen across browser tabs, reading prices, stock, delivery dates, returns terms and reviews, and narrows the list before the shopper looks at anything.

That moves some influence away from advertising and page design and towards the facts an agent can read and check. A product with no size chart, or a delivery date that only appears in a banner image, may simply drop off the shortlist.

Our forecast for the next two years is that more of the comparison step moves into assistants, especially for repeat and specification-led purchases, while shoppers keep the final decision on what they pay for.

Trust, privacy and control

Shoppers will only delegate if they trust what the agent does with their money and their data. We expect them to want the same controls a merchant wants: a spending limit, a confirmation step before payment, a clear record of what was bought and why, and an easy way back if something goes wrong.

For stores, that means being clear about how customer data shapes recommendations, asking for consent where the law requires it, checking that prices and offers do not treat groups of customers unfairly, and making returns as easy for an order an agent placed as for any other.

What to prepare now

You do not need a plan for every forecast in this guide. These steps pay off whichever way the next two years go:

See how agents see your store today. Work through the agent-ready store audit, or start with a free store audit.
Switch on what your platform already offers. If you are eligible for Shopify's Storefront MCP and Catalog MCP, or run a live BigCommerce store, check the servers are set up and returning accurate data.
Fix product data first: titles, variants, dimensions, materials, prices, stock, delivery and returns, consistent across product pages, feeds and structured data.
Write your approval rules before any agent acts: which changes need sign-off, who signs off, how each change is undone, and what is tested on a copy of production first.
Give each agent one job and one owner. Keep a shared log, and limit how much any single change can touch.
Keep WebMCP on your watch list while it sits behind a flag, and share WebMCP for ecommerce with your developers or agency.
Run voice and AR as small tests, each with one use, one measure and a date to review it.

Vortex IQ is the AI workforce for ecommerce. Specialist crews, such as Store Health (Pulse) and SEO and AI Visibility (Beacon), find problems in your own store data, prepare the exact change and check the result once it is applied. It proposes. Your team approves. Nothing goes live without your say-so.

Frequently asked questions

What is agentic commerce?

Buying and selling where AI agents do part of the work. On the shopper side, an agent finds, compares and sometimes buys products for a person. On the store side, agents check the store, prepare changes to content, prices or stock, and pass them to the team for approval.

Can AI agents already buy from online stores?

In part. Since 28 September 2026, Shopify has offered checkout tools for browser agents that run inside the checkout page, with no setup needed from the merchant. How far an agent gets elsewhere depends on the platform and on how clearly your pages and checkout read to software. The agent-ready store audit has the dates and details.

Is voice commerce worth planning for?

Plan for it through your data, not as a project of its own. Clear product names, plainly named variants, and delivery and returns information an assistant can repeat accurately help with voice, text assistants and search alike. If you sell repeat purchases, a small reorder test is a sensible place to start.

Is AR shopping worth investing in now?

Only if size, fit or how a product looks in a room decides the sale in your category, and then as a small test. Accurate dimensions and good photos come first, because every channel uses them.

Should AI agents act without approval once they know the store?

Not for anything that changes the live store, spends money or reaches a customer. Learning makes an agent's proposals better, but it can also learn the wrong lesson, and conditions change. Widen what an agent is trusted to prepare as its record of approved work builds up, and keep the sign-off for anything that goes live.

Where agentic commerce goes next: multi-agent systems, voice, AR and how shoppers will buy