Agentic Commerce

Visa, Mastercard, PayPal, OpenAI, Google, and Anthropic have all shipped products where an agent can research, compare, and buy. Six protocols are in production. Market prints for 2030-2031 run from $218B (Mordor Intelligence) to $3-5T (McKinsey). Those are not one market. 54% of US adults use AI agents. 24% trust them to make purchases.

Tap to enlarge

What is agentic commerce?

Agentic commerce is when AI agents handle shopping on your behalf. You tell an agent what you want, and it finds products, compares prices across stores, checks reviews, places the order using a tokenized payment method, and tracks delivery. You never open a browser or fill out a checkout form.

In traditional e-commerce, you browse, compare, and click. In agentic commerce, you say "get me running shoes under $150, good for flat feet, delivered by Friday" and the agent searches Nike, Amazon, and specialty retailers simultaneously, weighs price against reviews and shipping speed, and completes checkout. You go from doing the shopping to describing what you want.

You accept a job in a new city and need to relocate your family. Today, that means weeks of juggling real estate listings, moving company quotes, utility transfers, school research, and furniture shopping across dozens of websites. In an agentic world, you describe your budget, commute preferences, and kids' school requirements. The agent scans rental listings, cross-references commute times, identifies neighborhoods with the right schools, gets moving quotes, negotiates rates, schedules utility transfers, and coordinates furniture deliveries to arrive the day you move in. Weeks of browser tabs and phone calls become a few days of conversation.

The shift is structural. Each prior wave of commerce technology (web storefronts in the 2000s, mobile commerce in the 2010s, social commerce in the 2020s) added a new channel. Agentic commerce removes the human from the transaction loop. The agent is a new buyer. The interaction is machine-to-machine, governed by rules written into commerce protocols.

What shipped in 2026

Three things converged in 2025-2026. Large language models got good enough to parse shopping intent and make reasonable purchase decisions. Google, OpenAI, and Stripe shipped open commerce protocols so agents can talk to any merchant through a standard API. And payment infrastructure caught up: Visa, Mastercard, PayPal, Coinbase, and Stripe each built dedicated products for agent-initiated transactions, including crypto-native protocols like x402 that let machines pay machines. The adoption curve could be faster than prior platform shifts because agents traverse the same digital paths as human shoppers, riding on rails already laid by two decades of e-commerce.

What has already shipped

Every major platform and payment network has shipped agent commerce products in the past twelve months.

Tap to enlarge

Tap to enlarge

The protocol layer

The pace of protocol development is what separates this from prior hype cycles. Six interoperability standards are driving real adoption:

Model Context Protocol (MCP), created by Anthropic, standardizes how AI agents connect to external tools and data sources. Unlike one-off API integrations, MCP provides persistent, structured communication. An agent retains memory, reasoning, and objectives across environments. For commerce, this means a shopping agent can maintain context about your preferences, price sensitivity, and past purchases across every merchant it interacts with, rather than starting from scratch each time.

Agent-to-Agent Protocol (A2A), from Google, enables autonomous agents to coordinate and negotiate with each other directly, regardless of vendor or architecture. Your shopping agent communicates with a retailer's commerce agent through standardized JSON-RPC and HTTP. This supports long-running tasks (like finding the best price over several days), dynamic capability discovery (one agent learning what another can do), and multimodal collaboration. The practical result is that a buyer's agent and a seller's agent can reach agreement on price, delivery terms, and return policy in seconds, at machine speed.

Agent Payments Protocol (AP2), also from Google, creates cryptographically signed mandates that link user intent, shopping cart, and payment into a single auditable chain. When an agent makes a purchase, the intent ("buy this specific camera for under $800") is cryptographically bound to the transaction. This creates accountability and reduces dispute risk: if something goes wrong, the signed mandate shows exactly what was authorized.

Agentic Commerce Protocol (ACP), built by OpenAI and Stripe, powers in-chat purchasing in ChatGPT. It connects the conversation layer (where intent is expressed) directly to payment processing. The merchant receives a structured order through Stripe's infrastructure, and the user sees a purchase confirmation without leaving the chat.

x402 and MPP represent the crypto-native path. x402 operates at the HTTP layer: request a resource, get a 402 status code with payment details, send a signed stablecoin transaction in the HTTP header, receive the resource. MPP is the multi-rail version: stablecoins, card payments, and Lightning through one protocol. Both are designed for machines paying machines, with no human-oriented checkout flow.

Computer use agents (Anthropic's Claude, OpenAI's Operator) skip structured APIs. These agents control a mouse and keyboard to use existing websites built for humans, which matters where merchants have not built agent-ready APIs. The agent fills out forms, clicks buttons, and completes checkout the same way a person would, faster.

How a transaction works

A typical agentic purchase has six steps. You state what you want ("order two bags of Colombian coffee beans, same brand as last time"). The agent checks your purchase history, brand preferences, and budget. It searches multiple stores simultaneously and compares options by weighing price, shipping speed, reviews, and your past preferences.

When it finds the best match, it completes checkout through tokenized payment using Visa, Mastercard, or PayPal agent protocols. Your card details are never exposed to the agent. After purchase, it tracks shipping, handles returns if needed, and updates its model of your preferences for next time.

McKinsey identified three interaction models (mckinsey.com).

Tap to enlarge

In the agent-to-site model, your agent talks directly to a merchant's website or API. The merchant may not even know the buyer is an AI. In the agent-to-agent model, your shopping agent negotiates with a retailer's commerce agent, both communicating at machine speed through structured protocols. In the brokered model, an intermediary agent sits between you and multiple merchants, querying dozens of stores simultaneously.

Each purchase teaches the agent something. After three coffee orders, it knows your preferred roast, your price ceiling, and how fast you go through a bag. By the tenth order, it reorders before you run out. This is predictive commerce: the agent makes decisions based on patterns you never explicitly described.

The five layers

Agentic commerce runs on five layers, each depending on the ones above and below it.

Tap to enlarge

The Artemis Market Map

The startup map shows the same layering. According to the Agentic Commerce Market Map by Artemis Analytics, there are already over 173 companies operating across 14 distinct categories in the space.

Tap to enlarge

The market is currently heavily weighted toward foundational infrastructure rather than consumer-facing apps:

• Agent Frameworks & Tooling (33 companies): The largest category, building the core plumbing (e.g., MCP, A2A, thirdweb, Questflow, Privy).
• Payment Infrastructure (25 companies): Establishing the new rails (e.g., Stripe, Visa, Crossmint, Reveel, Superfluid).
• Facilitators & Discovery (37 companies): Helping agents find products and execute tasks (e.g., Coinbase, PayAI, Openx402, UCP).
• Wallets & Account Abstraction (32 companies): Managing agent identity and balances (e.g., Phantom, Dynamic, ERC-4337, Rhinestone).
• Data & Analytics (17 companies): Providing on-chain and off-chain insights (e.g., Dune, Nansen, Quicknode).
• Identity & Trust (17 companies): Verifying agents and managing permissions (e.g., Mastercard, AgentProof, Cred Protocol).

With only 7 companies primarily classified as "User Interfaces" (like ChatGPT and Claude), no consumer access point has won yet. The rails are being laid before consumer applications dominate.

Key numbers

These figures are not the same object. Mordor counts a retail-AI market. Morgan Stanley and Bain count US agent-driven e-commerce. McKinsey counts global transaction volume that an agent might touch. They are not one market.

36% of consumers say they are open to AI making purchases for them (SWAP Commerce), and 84% consider one-click checkout important (WPFactory). McKinsey projects agent-intermediated transaction volume could reach $3-5 trillion globally by 2030. During Cyber Week 2025, one in five orders involved an AI agent. Morgan Stanley estimates US e-commerce spending through agentic shoppers could reach $190-385 billion by 2030, representing 10-20% of total US e-commerce. Bain projects the US agentic commerce market at $300-500 billion by the same year.

What changes for consumers

Shopping becomes a conversation. You describe what you want, the agent handles everything from discovery through delivery. No browsing, no tabs, no checkout forms.

Brand loyalty weakens. If your agent switches you to a cheaper toothpaste that works as well, you probably will not notice. Predictive reordering kicks in: your agent knows you buy coffee every three weeks and handles it before you think about it.

Tap to enlarge

Tap to enlarge

Product pages built for human eyes matter less than structured data built for agent consumption: schema.org markup, real-time inventory feeds, machine-readable specs. Traditional SEO gives way to what Commercetools calls "Answer Engine Optimization" and "LLM Optimization" (commercetools.com). Display ads and retargeting lose relevance when the buyer is an algorithm. The traditional idea of visiting Amazon for shopping or Expedia for travel gives way to a single concierge that handles shopping and travel when your intent first surfaces.

What changes for businesses

Agentic commerce restructures how products get discovered, how buying decisions are made, and how customer relationships form.

Product discovery and engagement. Companies need to build agents that understand customer intent and proactively suggest products, bundles, or alternatives. Product catalogs need semantic and behavioral metadata embedded at the data layer, more than keywords on a webpage. When an agent asks "what running shoes do you have under $150 with good arch support in size 11?", the response needs to come from structured data.

Customer loyalty and personalization. Legacy loyalty programs (points, tiers, birthday coupons) break down when the customer never visits your app. Hyperpersonalized offers triggered by inferred intent replace generic promotions. If an agent knows a customer moves every two years based on lease history, furniture retailers can time outreach to coincide with the next likely move. The challenge is building persistent customer-context layers accessible by agents, and exposing loyalty eligibility through APIs rather than through sign-in walls.

Core commerce platforms. Existing e-commerce infrastructure was built around a human clicking through pages. Agents need structured transaction endpoints where they can query inventory, check pricing, and complete checkout in a single API call. Dynamic pricing, real-time inventory awareness, and machine-readable return policies become requirements.

Payments and fraud. Authentication moves from stopping bots to enabling the right agents to transact for the right customers. Know Your Customer (KYC) standards must extend to Know Your Agent (KYA): verifying users and the agents acting on their behalf. Tokenization, dispute resolution, and settlement systems need to carry richer context about agent identity within transaction metadata.

In-store operations. Physical retail does not get a pass. Agents need to synchronize digital and physical customer journeys: sharing context with store associates, accessing digitized store maps and inventory, and eventually using spatial computing for in-store navigation. The agent that helps you shop online should know what is available at the store nearest you and hold items for pickup.

Fulfillment and returns. Agents can automate fulfillment decisions, negotiate return logic, and coordinate post-purchase actions. This requires agent-ready fulfillment APIs and integration with multi-carrier and last-mile logistics through modular connectors. An agent that bought the wrong size should be able to initiate a return, select a replacement, and schedule delivery without the customer filing a support ticket.

A stack built for a human clicking pages will miss the customer that never loads a page.

How money flows change

Agentic commerce threatens existing revenue streams, particularly advertising. Retail media networks built on ad-based models face pressure as agents bypass traditional discovery channels. If your agent picks your toothpaste based on ingredient analysis and price comparison, the toothpaste brand's display ad budget is wasted.

Several new monetization models are emerging:

Bundled commerce and revenue sharing. Agents coordinate purchases across multiple brands, assembling them into coherent experiences. A honeymoon planning agent negotiates flights, hotels, excursions, and dining from separate providers. Each provider gets a share. The coordinating platform takes a service fee.

Real-time negotiation fees. Agents negotiate on behalf of users: seat upgrades, loyalty point redemptions, bulk discounts. Platforms enabling these negotiations charge success fees per transaction.

Premium agent capabilities. Third-party developers offer vertical AI agents (fashion stylists, travel planners, procurement specialists) through subscription plans or tiered access to advanced features.

Data insights. Brands pay for anonymized, agent-filtered consumer behavior analytics: which products get considered versus rejected, price sensitivities, competitor comparison patterns. This data is more granular than anything available through traditional analytics because agents evaluate products systematically rather than through emotional browsing.

Conversational marketplaces. AI agents evolve into full marketplaces where purchases happen through dialogue. Marketplace operators monetize through listing fees, commissions, and payment processing.

Inter-agent protocol fees. When agents from different platforms interact, monetization can happen through protocol-level fees for interoperability or commission sharing when joint value is created. This is analogous to interchange fees in the card network world, but applied to agent-to-agent communication.

Most agent interactions are still free, and providers have not developed clear revenue strategies.

The trust gap and fraud problem

There is a 30-point gap between agent usage and purchase trust: 54% of US adults use AI agents, but only 24% trust them to buy things. Top consumer concerns: payment security (32%), privacy (26%), agent errors (18%), loss of control (17%). According to a July 2025 Kearney US Consumer Study, roughly 30% of the market consists of "privacy-conscious skeptics" who demand absolute data control and assurances that agent mistakes will be financially corrected before they will adopt agentic shopping.

Tap to enlarge

Tap to enlarge

The fraud picture is worse than the trust numbers suggest. Visa's Fall 2025 report found that the threat model has shifted from stolen credit cards to compromising AI agents themselves. Malware that alters agent preferences to steer purchases toward fraudulent merchants. Fake stores designed to trick AI shopping agents. Traditional fraud detection was not built for the speed and scale of agent-driven attacks.

Legal grey zone

AI agents cannot form binding contracts in most jurisdictions. Visa's analysis notes that under GDPR and the ePrivacy Directive, agents capturing personal data via web scraping or mimicking tracking footprint cookies require explicit legal bases. Some laws also mandate direct user acceptance of terms and conditions, which leaves autonomous purchasing agents in a legal bind when they act without active confirmation.

Tap to enlarge

Systemic risk

Autonomous agents are decision-makers operating at scale. A single faulty prompt or model update can trigger cascading failures: incorrectly booked flights, overordered inventory, purchases made without proper consent. When agents are interconnected across multiple systems and protocols, minor errors can compound exponentially. The question is whether an agent makes good decisions today, and whether the system can contain and recover from failures. Can agents backtrack? Can purchases be reversed automatically?

Accountability

When an agent makes a bad purchase, who is liable? The platform that trained the model? The brand that deployed the agent? The user who set the spending limit? There is no global consensus. The EU AI Act provides partial clarity for high-risk systems, but enforcement is still evolving. In the US, fragmented regulations leave companies in a liability vacuum. Until frameworks solidify, companies are building their own guardrails: permissions layers, identity mapping, tiered trust levels, and explainable decision logs.

Data sovereignty

AI agents operate on data, and data sovereignty is increasingly a geopolitical issue. Countries like India and France enforce data localization. Europe is pushing for AI sovereignty. If an agent processes EU citizen data through a US-based API, compliance is unclear. If it is trained on global data but acts locally, the legal status is ambiguous. OpenAI has already launched a "for countries" model offering localized infrastructure and regulatory flexibility, which shows how hard it is to scale agent platforms without breaching national data protection and consumer protection laws.

Trust is contextual. What feels intuitive in one market can be unacceptable in another. For agents to earn trust, they need transparent decision trails, fail-safes, reversibility, and boundaries the user can set.

Competitive field

The core question: who controls the agent? That entity controls what gets bought.

Google. Open standard, owned distribution. UCP is open. Search and Gemini are not. Same playbook as Android.

OpenAI. Discovery and intent. About 800 million weekly active users. Moving from owning checkout to owning where the ask starts.

Coinbase. Machine-to-machine pay. If agents settle in USDC over x402, Coinbase takes a cut no matter which model wins the chat.

Shopify. Infrastructure on every AI surface, owned by none of them. Checkout Kit, Catalog, Universal Cart.

Stripe and Paradigm. One protocol across stablecoins, cards, and Lightning. Backed by Anthropic, OpenAI, Mastercard, and Shopify.

Amazon. Purchase history, reviews, and logistics behind Rufus and Alexa+. Agents can still read third-party reviews. That is the hole.