TL;DR
Agentic commerce is buying in which AI agents find products, compare offers and complete purchases for a person or business, within limits that person sets. It runs as a chain of five steps, from discovery and decisioning to checkout, payment and fulfilment. Open protocols split the work, with ACP and UCP handling checkout while AP2 and card-network tokens prove the buyer approved the payment. In 2026 OpenAI moved away from in-chat checkout, so merchants keep control of the final purchase. Merchants should start by cleaning product data, exposing live price and stock, and setting clear transaction limits. The safest first step is one small, approval-required workflow whose agent traffic you measure separately.
Key Takeaways Agentic commerce means AI agents discover, decide and complete purchases within limits a person or business sets. Protocols split the work, with ACP and UCP for checkout, AP2 and card-network tokens for payment authority, and MCP for tool access. OpenAI shifted from in-chat checkout to merchant-controlled checkout in 2026, so merchants keep the system of record. Readiness starts with clean product data, live price and stock, agent-usable APIs and explicit transaction limits. Liability for a wrong agent purchase is still unsettled, so every order needs an evidence chain of who approved what. Start with one bounded, approval-required workflow and measure agent traffic apart from human traffic. Watch on YouTube
Microsoft Fabric + Karl: Instant Retail Insights With AI
Watch Karl, Kanerika’s AI data analyst agent, turn retail questions into instant insights on Microsoft Fabric, the kind of seller-side agent that agentic commerce builds on.
When the Shopper Is Software
For thirty years retailers have studied shoppers, from their clicks to the second they hesitate over a shipping fee.
Then agentic commerce arrived and flipped that relationship. The visitor on a product page can now be software. It reads the feed instead of the photos, then checks stock through an API instead of scrolling.
That visitor never gets tired, ignores the banner ads, and answers to a human who set its budget. Visa reported in October 2025 that AI-driven traffic to US retail websites had grown more than 4,700% over the previous year. So the merchants who win this shift will be the ones whose data, prices and rules an agent can read and trust.
What Is Agentic Commerce?
Agentic commerce means an AI agent shops for a person or a company, from search to payment, inside limits a human approved. The agent does more than suggest a product. It can also build a cart, request a quote, apply a payment credential and confirm the order.
The money at stake explains why payment networks and AI labs moved so fast. McKinsey estimates agentic commerce could orchestrate as much as $1 trillion in US retail revenue by 2030. Globally, the estimate is $3 trillion to $5 trillion, according to Digital Commerce 360’s report on the study. Orchestrated means an agent shaped the purchase, which is a broader idea than an agent finishing every step alone.
How It Differs From E-Commerce, Chatbots and Agentic AI
Traditional e-commerce puts a person at every step, while a recommendation engine ranks products and still leaves the choice to the human.
A shopping chatbot answers questions, but it rarely touches the checkout. This model is what happens when agentic AI is allowed to act on a purchase. Also, as the comparison of agentic AI vs generative AI shows, an agent that acts needs different controls than a model that only writes.
The clearest way to compare these models is to ask who does each job and where the permission to pay lives.
Model Discovery by Decision by Checkout by Where authorization sits
Traditional e-commerce Shopper browses and searches Shopper Same person Card details entered at checkout Recommendation engine Algorithm ranks options Shopper Same person Card details entered at checkout Shopping chatbot Chatbot suggests products Shopper Person on the merchant site Card details entered at checkout Rule-based replenishment Fixed rule, such as reorder at a threshold The rule System Stored payment plus a preset rule Agent-assisted commerce AI agent Shopper approves the cart Agent prepares, shopper confirms Approval at each purchase Agent-executed commerce AI agent Within limits set in advance The agent Mandate or scoped token set in advance B2B procurement agent Approved suppliers, searched by the agent Purchasing policy Agent raises the purchase order Contract terms, cost center and approval rules
Three Levels of Autonomy: Agent-Assisted, Agent-Executed, Agent-to-Agent
Autonomy works like a dial with three settings. At the agent-assisted level, the agent researches and prepares a cart while the shopper presses the final button.
At the agent-executed level, the agent buys on its own inside rules set in advance. For instance, it reorders printer paper when stock drops, as long as the price stays under $40.
The third level is agent-to-agent commerce, where a buyer’s agent negotiates with a seller’s agent over price, delivery or terms. It is also the least mature setting, and accountability is hardest to trace there.
An agent’s technical ability to pay says nothing about whether it has permission to pay. That gap between capability and authority is where most of the design work sits, and it shapes every step of the journey below.
How Agentic Commerce Works: From Discovery to Transaction
Every agent-led purchase moves through the same five steps, even though products differ. Discovery and decisioning happen on the agent’s side, while checkout, payment and fulfilment depend on systems the merchant and the payment provider control. A consent and audit trail has to run underneath all five, because a dispute months later will ask who approved each step.
Discovery: How Agents Find Products
An agent finds products by reading structured data rather than scanning pages the way a person does. It pulls product feeds, schema markup, merchant APIs and search results, then filters them against the shopper’s constraints. Agentic browsers handle the sites that have no API, although clicking through a page built for humans is slower and more fragile.
Discovery is also where merchants first win or lose. A catalog with missing sizes, vague titles or stale prices drops out of the shortlist, and the shopper never sees those products.
Decisioning: How AI Agents Make Purchasing Decisions
AI agents make purchasing decisions by scoring each eligible offer against the instructions they were given. Price, delivery date, return policy, brand preferences and past orders all feed the comparison, together with any budget cap.
A good agent also explains its pick. So the human can see why it chose the $89 option over a $79 one that ships in two weeks.
Weak decisioning usually comes from weak inputs, such as messy product titles. When two retailers describe the same product differently, the agent may treat them as different items. Consistent attributes therefore matter more than clever ranking logic.
Checkout: Turning Intent Into an Order
Checkout is where intent becomes a binding order, and it needs safeguards that human checkouts rarely think about. Quotes need an expiry time, because prices change between the moment an agent compares and the moment it buys. Each order request also needs an idempotency key, so a retried call after a network timeout does not create a duplicate order.
For this reason, many teams separate a prepared checkout from an authorized one. The agent assembles the cart, shipping and total, and then a human or a pre-approved mandate releases it. That split keeps the merchant’s order system, connected through AI agent integration with the ERP, OMS and product catalog, as the single source of truth.
Payment: Identity, Delegated Authority and Tokens
Payment asks two separate questions. Identity asks whether this agent is who it says it is, which is the job of AI agent authentication . Delegated authority asks whether the human really allowed this purchase, at this amount, from this merchant.
Payment tokens then answer the second question without handing the agent a raw card number. A scoped token can be limited to one merchant and one cart total, so a compromised agent cannot reuse it elsewhere.
Fulfilment and Post-Purchase Support
Fulfilment looks familiar once the order exists, yet agents change what happens after it. For example, a buyer’s agent may track delivery, file a return or open a support case on the shopper’s behalf.
Merchants that already run contact center automation answer those agent requests faster. That is because the same order and case data serves both humans and software.
Accurate stock data matters here too, because an agent that bought an item shown as available will escalate fast when it ships late. Teams that use AI inventory management to keep availability current avoid much of that friction.
Kanerika Service
Agentic AI Services
Kanerika designs and builds AI agents that connect to your catalog, order and payment systems with clear limits on what they may do.
Explore Agentic AI Services Agentic Commerce Protocols: ACP, UCP, AP2, MCP and the Card Networks
Several open protocols now share the work of agent-led buying, and each one covers a single layer of the purchase. A merchant usually meets more than one, because checkout, payment authority and tool access are separate problems. So knowing which layer each protocol owns saves a lot of confused vendor meetings.
Agentic Commerce Protocol (OpenAI and Stripe) and Shared Payment Tokens
Stripe released the Agentic Commerce Protocol (ACP) on September 29, 2025, as an open standard codeveloped with OpenAI . The ACP specification is open source under Apache-2.0 and is maintained by OpenAI and Stripe. In practice, it defines how a buyer’s agent and a business exchange product, cart and order details.
The payment piece is the Shared Payment Token (SPT). Stripe issues it so that an app like ChatGPT can start a payment without seeing the buyer’s card details.
Each token is also scoped to one merchant and one cart total. A leaked token is far less useful to an attacker.
Universal Commerce Protocol (Google With Shopify, Walmart and Target)
Google published the Universal Commerce Protocol (UCP) on January 11, 2026, as an open-source standard for agent-led shopping. Google developed it with Shopify, Etsy, Wayfair, Target and Walmart.
More than 20 partners also endorse it, including Adyen, American Express, Best Buy, Mastercard, Stripe, The Home Depot and Visa. UCP works with AP2 and connects through APIs, A2A or MCP.
One design choice matters most to retailers. Under UCP the merchant owns its business logic and remains the merchant of record. Meanwhile, the shopper may start in AI Mode in Google Search or in the Gemini app.
Agent Payments Protocol (AP2): Mandates as Proof of Intent
Google announced the Agent Payments Protocol (AP2) on September 16, 2025. More than 60 organizations joined, among them Adyen, American Express, Coinbase, Etsy, Mastercard, PayPal and Salesforce.
AP2 is payment-agnostic, working across cards, bank transfers and other methods, and it can be used as an extension of A2A and MCP. Its core job is proof: it records what the user approved, while existing payment methods carry the money.
That proof comes as mandates, which are tamper-proof digital contracts signed with verifiable credentials. When the shopper is present, an Intent Mandate captures the request and the shopper’s approval signs a Cart Mandate. When the shopper is absent, they sign a detailed Intent Mandate upfront with price limits, timing and other conditions.
Model Context Protocol: Tool Access, Not Payment Authority
Anthropic introduced the Model Context Protocol (MCP) in November 2024 so that agents had a standard way to call tools and read data. In December 2025 Anthropic donated MCP to the Agentic AI Foundation , a fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. In commerce, MCP lets an agent query a catalog or an order API, yet it grants no permission to spend money.
Visa Intelligent Commerce, Trusted Agent Protocol and Mastercard Agent Pay
The card networks, by comparison, attack the problem from the payment side. Visa introduced Visa Intelligent Commerce on April 30, 2025, opening its network to developers building agents. Its launch partners included Anthropic, IBM, Microsoft, Mistral AI, OpenAI and Perplexity, as well as Samsung and Stripe. In Visa’s words, “Each consumer sets the limits, and Visa helps manage the rest.”
Visa followed with the Trusted Agent Protocol on October 14, 2025, built with Cloudflare to help merchants tell legitimate agents from malicious bots. Mastercard launched Agent Pay on April 29, 2025, with Agentic Tokens that build on the tokenization behind mobile contactless payments. Banks and issuers weighing these programs will find more on the risk side in our guide to agentic AI in financial services .
What Changed in 2026: Instant Checkout Gives Way to Merchant-Controlled Checkout
OpenAI first used ACP for Instant Checkout in ChatGPT , launched in September 2025. US users could then buy single items from US Etsy sellers in chat.
During 2026 OpenAI shifted away from that model. Checkout.com reported in April 2026 that AI now handles discovery and intent while merchants keep control of checkout. ACP is also expanding into product discovery.
The lesson for merchants, then, is practical. Plan for agents that send you qualified, ready-to-buy shoppers, and keep your own checkout, order records and customer relationship strong enough to receive them.
Protocol and Payment Components at a Glance
The table below sums up which layer each component owns and what it leaves to others.
Component Job it does What it does not do Maintained by Status (November 2026)
Agentic Commerce Protocol (ACP) Agent-to-merchant exchange of product, cart and order data Prove what the buyer approved OpenAI and Stripe Open spec under Apache-2.0, expanding into discovery Shared Payment Token (SPT) Starts a payment without exposing card details Work beyond its merchant and cart total Stripe Announced with ACP, September 2025 Universal Commerce Protocol (UCP) Common way for agents to shop with merchants Replace the merchant of record Google, with Shopify, Etsy, Wayfair, Target and Walmart Published January 2026, works with AP2 Agent Payments Protocol (AP2) Signed mandates as proof of user intent Replace payment rails; it works across payment methods Google, with 60+ organizations Announced September 2025 Model Context Protocol (MCP) Connects agents to tools and data Grant permission to pay Agentic AI Foundation (Linux Foundation) Donated December 2025 Visa Intelligent Commerce and Trusted Agent Protocol Agent-ready credentials and agent verification for merchants Choose what the agent buys Visa, with Cloudflare for TAP Announced April and October 2025 Mastercard Agent Pay Agentic Tokens for agent-initiated card payments Check product data quality Mastercard Announced April 2025
In short, none of these components replaces the others. UCP is documented as compatible with AP2. Still, no public documentation shows ACP and UCP working together out of the box, so plan for more than one integration.
AI Assessment
How Ready Is Your Business for AI Agents?
Score your data, AI foundations and agent readiness in a few minutes and get recommendations for your next step.
Start Your AI Assessment → Closed vs. Open Models: Who Owns the Customer?
Agent-led buying is splitting into two models, and the choice affects margin and customer data more than technology. In a closed model, one platform runs the catalog, the checkout and the customer record inside its own walls. The merchant gets reach, but the platform owns the relationship and sets the fees.
In an open, protocol-based model, the agent talks to the merchant through a shared standard. The merchant stays the merchant of record, keeps the order data and can serve agents from several AI platforms with one integration. UCP, for instance, is built around that arrangement.
Most retailers will live with both for years, just as they sell through marketplaces and their own site today. The AI agent marketplace trend points the same way, since buyers will reach products through many agents at once. A sensible rule is to accept closed channels for reach while keeping product data, pricing logic and order history in systems you control.
Use Cases and Examples: Where AI Agents Already Shop
AI agents show up first where purchases are repetitive, comparable or time-sensitive. The examples below mix illustrative scenarios with the deployments that are documented so far.
Consumer Shopping: Electronics, Household Replenishment and Travel
Three illustrative scenarios show the pattern. A shopper asks an agent for noise-cancelling headphones under $250 with next-week delivery, and the agent returns a ranked cart for approval. Another sets a standing rule to reorder detergent and coffee when stock runs low, within a monthly cap.
Travel is the third case, because the agent can rebook a cancelled train within a preset fare limit and send the receipt afterward. Still, all three depend on clean product data and on limits the shopper can see and change.
Documented Deployments So Far
Etsy sellers in ChatGPT were an early large public example, through Instant Checkout in September 2025, before OpenAI moved to merchant-controlled checkout in 2026. Google names AI Mode in Google Search and the Gemini app as the consumer surfaces where UCP lets merchants show and sell their products. We found no independently verified conversion or merchant adoption figures for any of these protocols, so treat vendor claims with care.
Merchant-Facing Agents: Pricing, Inventory and Catalog
Some of the most useful agents work for the seller rather than the shopper. They adjust prices within guardrails, flag stock that will run out before a launch, and fix catalog gaps that would hide products from buyer agents. Brands also build their own assistants, and a custom GPT trained on a product catalog is a common first step.
These seller-side agents rely on the same retail data foundations as buyer-facing commerce. Kanerika’s own retail work on pricing and forecasting, covered later, sits in this layer.
Case Study
AI-Powered Dynamic Pricing for a Global Fashion Retailer
Kanerika built AI pricing with 100% auditability of pricing decisions, cutting price change cycle time by 39%.
Read the Case Study → B2B Buying: Procurement and Reorder Agents
B2B may be where agent-led buying grows largest. Gartner forecasts that AI agents will intermediate more than $15 trillion in B2B spending by 2028, as reported by Digital Commerce 360 . Business purchases already run on rules, budgets and approvals, which gives agents a clearer frame than most consumer shopping.
Approved Replenishment and Indirect Spend
Indirect spend is the natural starting point, since office supplies, spare parts and software seats are low-risk and highly repetitive. An agent can reorder from approved suppliers at contract prices, charge the right cost center and stop when a budget line runs out. In effect, the approved supplier list and contract terms become the agent’s guardrails.
Quotes, Purchase Orders and Invoice Matching
For larger buys, the agent collects quotes, compares them on price and lead time, and drafts a purchase order for approval. Once goods arrive, it matches the invoice against the order and the receipt. That flow only works when procurement, ERP and supplier systems talk to each other, which is the job of enterprise application integration .
Agent-to-Agent Buying With Human Accountability
Segregation of duties still applies when software does the buying. The agent that raises a purchase order should not be the one that approves payment.
Also, a supplier’s request to change bank details must always go to a person. That one rule closes a well-known route for invoice fraud.
A human-in-the-loop design for AI agents keeps a named person accountable for each approval tier. Agents move the routine work, while people own the exceptions and the final sign-off on anything unusual.
What Shoppers and Merchants Gain From AI Agents
The gains show up on both sides of the purchase, but only when the data underneath is sound. Still, each group gets something different.
For Shoppers and Business Buyers
Shoppers get time back, because an agent compares more offers than a person would and remembers their preferences. Dull reorders also disappear from the to-do list. For business buyers, an agent that applies contract prices and budget rules every time cuts off-policy spend.
For Merchants and Brands
Merchants gain visitors who arrive with clear intent and a budget already set. Visa’s 4,700% jump in AI-driven traffic to US retail sites shows that shoppers already arrive through AI, even before agents complete the purchase. Brands with clean data also get compared fairly on price, delivery and returns, rather than on ad spend alone.
When the Benefits Do Not Show Up
The benefits fail to appear when the basics are weak. Agents skip products with thin data, abandon checkouts that demand a human login, and stop trusting merchants whose stock figures prove wrong. In each case the fix is the same groundwork described in the readiness checklist below.
Agentic Commerce Risks: Fraud, Liability and Brand Control
Every risk here comes back to one question, which is whether a specific human approved this specific purchase. So the controls below each protect a different part of that answer.
Identity Is Not Authority
A verified agent can still make an unapproved purchase, since identity proves who is acting while authority proves what they may do. Keep the two checks separate, and set per-merchant, per-amount and per-category limits through a clear permission control framework for agentic AI .
Prompt Injection and Untrusted Product Content
Product pages and reviews are untrusted input. Hidden text that tells an agent to ignore its budget is a real attack path.
So AI agent guardrails must treat page content as data rather than instructions. The wider threat model sits in our guide to agentic AI security .
Who Pays When an Agent Buys the Wrong Thing? Chargebacks and Disputes
Nobody has settled liability for a wrong agent purchase yet. The outcome depends on the payment method, the contracts between platform, merchant and shopper, and card network rules. The best protection is evidence, meaning a signed record of intent, the approved cart, a scoped token and a matching order log.
That chain also helps fraud teams, who can use AI in fraud detection to spot agent behavior that breaks from the approved pattern.
Privacy, Disclosure and Brand Visibility
Consumer protection and privacy rules still apply to agent purchases. Disclose clearly when an agent acts, get consent before storing payment credentials, and keep only the data the purchase needs, with counsel checking the details.
Brand risk is quieter, because a shopper who never visits your site never sees your story. So product data has to carry the brand on its own.
How to Prepare for Agentic Commerce: A Merchant Readiness Checklist
Readiness for agentic commerce is mostly data and plumbing work that already pays off for human shoppers. Each step below builds on the one before it, so start at the bottom and climb.
Fix Product Data and Feeds
Complete attributes, consistent units, accurate images and current schema markup are the price of entry. Structured data helps an agent understand a product, although it cannot guarantee the agent picks it. Data governance services give each attribute an owner, so quality holds after the first cleanup.
In practice, that means Product markup with a stable sku and gtin, plus an Offer carrying price, priceCurrency and availability. Google’s merchant listing documentation also covers shippingDetails and hasMerchantReturnPolicy, so delivery and return answers come from your data. Generate the markup from the system that runs checkout, so it never quotes a stale price.
Make Storefront and APIs Usable by Agents
Expose live price, stock and delivery estimates through APIs, and let verified agents check out without a captcha built for humans. Plan the operating model alongside the tech, using our agentic AI in retail strategy guide.
Agent verification is moving to signed requests. Visa’s Trusted Agent Protocol has approved agents send agent-specific cryptographic signatures, built on the HTTP Message Signature standard. That lets a merchant open the checkout path to signed agents while bot rules keep blocking unsigned traffic.
Set Transaction Boundaries and Measure Agent Traffic Separately
Define what an agent may buy, from whom, how much and how often, then decide which orders need a human. Tag agent sessions apart from human ones, since blended numbers hide whether agents convert, return items or fail at checkout. AI agent observability , in particular, gives you that separate view.
Write the limits down as a policy each agent carries: a per-order ceiling, a monthly cap, approved suppliers, allowed categories and an expiry date. Those fields mirror the price limits and timing an AP2 Intent Mandate records. The rule the buyer set then travels with the payment as signed proof.
Adjust SEO and AEO for AI Answers
Discovery now happens inside AI answers as well as search results. Track how often AI assistants mention and cite your products, and write product content that answers the questions buyers ask agents.
Capability Minimum evidence Common gap KPI to track
Product data and feeds Full attributes and schema on top sellers Missing sizes, units or variants Share of SKUs with complete attributes Live price and stock API returns current price and availability Feeds refreshed once a day Price and stock mismatch rate Agent-usable checkout Verified agents can complete an order Captcha or forced account login Authorized-checkout completion rate Transaction limits Written limits per agent, amount and category Limits held only in vendor settings Unauthorized attempts blocked Agent traffic measurement Agent sessions tagged in analytics Agents counted as human visitors Agent conversion and return rate
A Rollout Roadmap With Human Checkpoints
A phased roadmap keeps early mistakes small, so they stay visible and cheap to fix. Each phase ends with a human checkpoint before the agent gets more authority.
The first phase picks one workflow, such as reordering office supplies, and records today’s cost and error rate.
Next, phase two designs authority limits and connects order, catalog and payment systems, often with AI agent orchestration when several agents share the work. After that, a read-only or approval-required pilot runs in phase three , and phase four allows bounded transactions before scaling.
Checklist
Agentic AI Checklist
A step-by-step checklist for scoping, governing and piloting AI agents before they touch real transactions.
Get the Checklist → During the pilot, review agent decisions with a structured method like LLM-as-a-judge together with human spot checks, as our guide to AI agent evaluation describes. Track authorized-checkout completion, duplicate orders, blocked unauthorized attempts, disputes and refunds, and margin after returns.
How Kanerika Helps Merchants Get Agent-Ready
Kanerika works on the layers agentic commerce depends on, which are clean retail data, governed AI decisions and integrated order systems. Our agentic AI services follow a simple path to assess readiness, design agent authority, build integrations, govern decisions and train teams. That work also builds on AI and ML services and on our experience in retail and FMCG .
Three Kanerika retail projects show that groundwork, though none of them is an agentic checkout deployment. A dynamic pricing engagement for a global fashion retailer cut price change cycle time by 39% and raised margins on top SKUs by 24%. An AI clienteling project for a luxury brand made client preparation 48% faster and lifted transaction value by 33%.
For a luxury fashion house, AI demand forecasting cut inventory holding costs by 37% and improved forecast accuracy by 87%. In other words, auditable pricing, accurate stock and connected customer data are what buyer agents will test first.
Case Study
AI Demand Forecasting for a Luxury Fashion House
Kanerika’s forecasting work cut inventory holding costs by 37% and lowered stockouts during launches by 22%.
Read the Case Study → What Comes Next for Agent-Led Buying
Three developments to 2028 look plausible, and they are scenarios rather than forecasts. First, agent-assisted discovery grows fastest, because it needs no new payment authority. Second, bounded replenishment and B2B procurement scale next, since rules and budgets already exist there. Third, open interoperability across protocols arrives only if authority and dispute rules mature.
Agentic commerce is AI agents buying within human-set limits, and it works through discovery, decisioning, checkout, payment and fulfilment. For merchants and buyers, the move now is to clean the data, expose live systems, write down the limits and pilot one approval-required workflow.
Frequently Asked Questions
What is agentic commerce?
Agentic commerce is buying and selling in which AI agents act for a person or a business. The agent searches for products, compares offers, builds a cart and can complete payment, but only inside limits a human has approved. Those limits cover amount, merchant, product category and when the agent must ask before buying.
How does agentic commerce work?
It works in five steps. The agent discovers products through feeds and APIs, scores offers against the buyer’s rules, prepares a checkout, and pays with a scoped token or signed mandate. The merchant then fulfils the order. A consent and audit trail runs underneath every step, so each purchase can be traced back to an approval.
How is agentic commerce different from traditional e-commerce?
In traditional e-commerce a person browses, decides and enters card details at checkout. In agentic commerce software does the browsing and comparing, and it may also complete checkout. The merchant’s job shifts from persuading a human on a page to giving an agent accurate, machine-readable data, live prices and an agent-friendly checkout.
What is the Agentic Commerce Protocol?
The Agentic Commerce Protocol, or ACP, is an open standard that Stripe and OpenAI released in September 2025. It defines how a buyer’s agent and a business exchange product, cart and order details. Stripe’s Shared Payment Token lets the agent start a payment without seeing card details, and each token is limited to one merchant and cart total.
What are agentic payments?
Agentic payments are payments an AI agent starts on a person’s behalf. They rely on tokens and signed approvals instead of raw card numbers. Examples include Stripe’s Shared Payment Tokens, Mastercard Agentic Tokens, Visa Intelligent Commerce credentials and Google’s AP2 mandates, which record exactly what the buyer approved before money moves.
What is the difference between ACP, UCP, AP2 and MCP?
ACP and UCP both define how agents shop and check out with merchants, with ACP from OpenAI and Stripe and UCP from Google and retail partners. AP2 records proof that a user approved a purchase. MCP connects agents to tools and data but grants no permission to pay. Most merchants will meet more than one.
Can AI agents buy without human approval?
Technically yes, if a person signs a mandate or sets limits in advance, for example a standing rule to reorder supplies under a fixed price. Most merchants and buyers still start with approval-required purchases, where the agent prepares the cart and a person confirms it. Autonomy should grow only after the pilot data supports it.
Who is liable if an AI agent makes the wrong purchase?
Liability is still unsettled. It depends on the payment method, the contracts between the AI platform, the merchant and the shopper, and card network rules. The practical defense is evidence, meaning a signed record of intent, the approved cart, a scoped payment token and an order log that ties them together for any dispute.
How can retailers prepare for agentic commerce?
Retailers should complete product attributes and schema markup, expose live price and stock through APIs, and make checkout usable by verified agents. They should also write down transaction limits, tag agent traffic separately in analytics, and pilot one approval-required workflow before expanding. Each step also helps human shoppers, so the work pays off either way.
Does agentic commerce apply to B2B procurement?
Yes, and B2B may grow faster than consumer shopping because purchases already follow rules, budgets and approvals. Gartner forecasts AI agents will intermediate more than $15 trillion in B2B spending by 2028. Good starting points are indirect spend, approved replenishment, quote comparison and invoice matching, with people approving exceptions such as supplier bank-detail changes.
