What is agentic commerce?

Agentic commerce is a model of online buying in which AI agents act on a shopper's behalf — interpreting a need, researching and comparing options, and completing the purchase and its follow-up, within limits the shopper sets in advance. Instead of a person searching, filtering, clicking, and checking out manually, an agent carries out the work: it understands the request, assembles a plan, and executes the transaction through connected systems once it has the shopper's go-ahead.

The term pairs two ideas. "Agentic" describes the autonomy — these systems pursue a goal and sequence the steps to reach it, rather than waiting for instruction at every click. "Commerce" describes the domain — the goal is a completed, satisfying purchase and the relationship around it, not just a recommendation. The result is software that behaves less like a search box and more like a capable personal shopper with permission to act.

This page covers what agentic commerce is, how it works, how it differs from traditional ecommerce and from chatbots, the protocols emerging underneath it, where it meets customer service, what it takes to work well, and where it falls short.

Agentic commerce in one sentence

Agentic commerce is AI that completes the purchase, not just suggests it.

How agentic commerce works

Most descriptions of agentic commerce reduce to a three-step loop. An agent moves through these stages in a single request rather than handing the work back to the shopper between steps.

Understand the intent. The agent interprets a request in full context. "I need durable trail-running shoes under $150, delivered by Friday" carries constraints on use case, price, and timing all at once — not three separate keyword searches.

Reason and plan. The agent assembles an approach: which retailers to check, which reviews or specs to weigh, and which real-time signals — inventory, pricing, shipping windows — matter to the decision.

Act. This is the defining stage. Through APIs and emerging commerce protocols, the agent closes the loop: it builds the cart, applies the shopper's stored preferences and credentials, and processes payment once the shopper approves. The shopper manages the agent rather than the store.

The capabilities underneath are familiar AI building blocks: large language models for understanding and reasoning, structured product data so the agent reads catalog details accurately, and secure connections to the retailer systems where orders, payments, and fulfillment actually live.

How agentic commerce differs from what came before

Traditional ecommerce AI is passive. A recommendation engine says "you might also like this," but the shopper is still the engine — searching, comparing, reading reviews, and clicking through checkout. Success is measured in clicks and sessions.

Agentic commerce inverts that. The shopper states a goal and the agent does the work, from discovery through payment. The difference is action: a chatbot or recommender describes and suggests; an agent executes.

The shift also changes how brands get found. In the traditional model, brands optimize content and keywords so a person clicks a link. In the agentic model, brands optimize structured, accurate product data so an agent can trust the listing enough to buy it — a move some describe as the turn from search optimization toward optimizing for AI-driven discovery. Visibility starts to depend on data quality and reputation rather than on headlines.

This is happening at real scale. Adobe Analytics reported that traffic to U.S. retail sites from generative AI sources grew roughly 690 percent year over year during the 2025 holiday season, and that shoppers arriving from AI assistants browsed longer and bounced less — signals of higher intent. Estimates of where this lands by 2030 vary widely with definition: Bain projects $300 to $500 billion in U.S. agentic commerce, while broader analyses of agent-mediated transactions run into the trillions globally. The range reflects how new the category is, not disagreement about its direction.

The protocols that make agentic commerce work

For an agent to act across many brands, it needs a standard way to connect. Several protocols are emerging, and they do different jobs. A brand generally needs more than one, and none has clearly won.

MCP (Model Context Protocol), from Anthropic, is a universal integration standard — the common way an AI agent discovers and calls a brand's tools, like order lookup, product search, or return processing. Every major AI provider has adopted it, which makes it the practical starting point.

ACP (Agentic Commerce Protocol), from OpenAI and Stripe, is a commerce-specific standard defining how product feeds, checkout sessions, and delegated payment should be structured so an agent can transact with a brand.

UCP (Universal Commerce Protocol), from Google and co-built with Shopify, is a competing commerce standard covering cart, real-time catalog, and shopper identity, with live checkout on Google's AI surfaces.

A2A (agent-to-agent), from Google, defines how one agent delegates a task to another — the layer that matters as shoppers begin using their own personal agents that talk to a brand's agent.

AP2 (Agent Payments Protocol), from Google, addresses secure payment authorization, using signed mandates that record a shopper's intent for a specific transaction.

The takeaway for brands is sequencing, not betting on one winner. For a fuller breakdown of how these relate and where to invest first, see the Gladly guide to agentic commerce protocols.

Where agentic commerce meets customer service

Most coverage frames agentic commerce as a shopping assistant — an agent that finds and buys products. That is the most visible layer, but it is not the whole picture. The larger shift is that the entire commerce experience is becoming conversational, and buying and service are converging into one continuous relationship.

Customers have always seen it this way. Someone who messages a brand about a return and, in the same breath, asks about a new product is not switching from "support mode" to "shopping mode." To them, both are just talking to the brand. The systems behind the curtain have been fragmented — a search tool, a chat widget, a returns portal, an order tracker — but the customer's mental model never was.

That convergence is why agentic commerce matters to customer service teams specifically. A conversation about a return is also a moment to offer a better-fitting exchange. A question about order status is a moment of engagement. The same agent can potentially recognize a natural commerce moment and act on it the way a good store associate would — not as a sales pitch, but as genuinely helpful context. Done well, the same conversation handles efficiency and revenue, resolution and relationship, at once. For more on this, see the Gladly perspective on what agentic commerce means beyond shopping agents and how it reshapes customer loyalty.

What agentic commerce needs to work well

Agentic commerce depends on something most retail stacks do not have by default: a complete, connected picture of the customer and the catalog.

For an agent to act on a shopper's behalf — and to recognize that a return is also an exchange opportunity, or that this shopper buys for function over fashion — it needs more than the words in the current session. It needs the relationship: past purchases, sizing learned from a previous exchange, channel history, and stated preferences, organized around the person rather than scattered across disconnected cases. When that context lives in one place, the agent already knows who it is talking to; when it does not, the agent rebuilds understanding from scratch every time, no matter how capable the underlying model is.

The same is true of product data. Agents are unforgiving about accuracy — outdated inventory or inconsistent catalog details cause a transaction to fail, and an agent that fails once tends to route around that retailer next time. Clean, structured, real-time data is a prerequisite, not a nice-to-have.

This is why agentic commerce is often described as an architecture problem rather than an intelligence problem. The agent's usefulness compounds when every conversation, purchase, and channel feeds one continuous record — and that compounding context is what turns a competent one-off interaction into a relationship a customer comes back to.

Strengths of agentic commerce

It removes friction from buying. Agents handle the tedious parts — comparison, forms, checkout, reorders — so a purchase that took a dozen steps takes one approval. For routine and repeat buying, that convenience is the entire value.

It personalizes at a scale that used to be reserved for a few. An agent that remembers sizing, preferences, and past context can deliver the kind of attentive service a personal shopper offers, applied to every customer rather than a handful.

It turns service conversations into commerce moments. Because the same agent can resolve an issue and recognize a natural next step, brands can earn revenue inside conversations they are already having — without bolting a separate sales motion onto support.

It compounds. Every post-sale interaction teaches the agent something that improves the next pre-sale moment. A return informs the next fit recommendation; a product comparison informs the next set of options. The experience gets better with use, which is what sustains loyalty without leaning on discounts.

Limitations of agentic commerce

It runs on trust that is easy to lose. Handing an agent access to payment details and personal data requires confidence that it will act in the shopper's interest and keep that data secure. A single mishandled transaction or unauthorized charge does lasting damage, and customers judge agent failures harder than human ones.

It is only as good as the data underneath it. An agent on top of fragmented customer records or stale catalog data produces agentic-sounding behavior without agentic capability — generic recommendations, failed checkouts, and missed context. Much of the early market is still here.

Discovery on third-party surfaces is not the same as owning the relationship. Showing up inside ChatGPT, Gemini, or Perplexity is real and growing, but those surfaces control discoverability and take margin over time, as every prior commerce channel has. Early experiments bear this out — some in-chat checkout features converted markedly worse than a brand's own site and were pulled back. Discovery through AI surfaces works; the brand that controls the experience tends to control the conversion.

It is not appropriate for every decision. High-stakes, sensitive, or judgment-heavy purchases — and the moments around them where something goes wrong — still benefit from human involvement and clear paths to reach a person. Designing for where the agent should hand off is as important as designing for what it can do alone.