Agentic commerce: AI agents now buy for customers

Customers increasingly do not open ten browser tabs. They ask an AI assistant to find a product, compare offers, and pick the best one, and in some channels the agent completes the payment too. This model is called agentic commerce, and in 2026 it moved out of the pilot phase: on 6 July, Salesforce made its Agentforce Commerce agents generally available, while Google and the payment networks are building the infrastructure underneath. The rule of the game changes for companies. A machine decides on the product first, and a human sees it second. A company that is unreadable to the agent drops off the shortlist before the customer ever looks at the offer.

What happened

On 6 July 2026, Salesforce announced the general availability of Agentforce Commerce, a set of sales agents: the Shopper Agent guides a customer from search to cart, the Buyer Agent handles purchases, and the Merchant Agent works on the seller's side. Salesforce said these agents integrate natively with ChatGPT and Google Search, including AI Mode, as well as the Gemini app. This is one of the first cases where agents that carry a customer through the full path to purchase ship as a finished product rather than an experiment.

The infrastructure underneath has been forming for more than a year. In September 2025, OpenAI and Stripe launched Instant Checkout in ChatGPT, built on the open Agentic Commerce Protocol. Catalogs of merchants on Shopify and Etsy were integrated with no extra setup, and the fee is paid by the merchant, with no effect on the price for the buyer and no effect on the ranking of results. Google took its own route and built Universal Cart plus the AP2 payments protocol, which lets an agent prove to a merchant or payment network that a real person authorized a given purchase. Google released AP2 as an open standard and contributed it to the FIDO Alliance.

The payment networks are part of the picture too. As CNBC reported, Visa and Mastercard are building agentic tokens that cryptographically confirm an authorized agent is on the other side of a transaction rather than a malicious bot. The scale is real: McKinsey estimates agentic commerce could redirect 3 to 5 trillion dollars of global retail spend by 2030, and the US B2C retail market alone up to 1 trillion dollars of agent-orchestrated revenue.

Keep the proportions in mind. Gartner places agentic AI at the peak of inflated expectations and reports that only 17 percent of organizations have deployed agents, though more than 60 percent plan to within two years. Liability and dispute rules for agent purchases, as CNBC notes, are still unresolved. In other words, discovery through AI assistants is running ahead of transaction completion, and companies have time to prepare, though less than it seems.

Why it matters for business

Agentic commerce moves the moment of decision. Until now a company fought for a human's attention: rank in search, the click, the look of the product page. When the first recipient of the offer is an agent, what counts is whether it can read and understand the company's data. A description written for a person does not help if price, availability, and product attributes are not available in a form a machine will accept without guessing.

Who controls the customer relationship also shifts. When a purchase goes through ChatGPT, Gemini, or a seller's agent, the platform holds the conversation context and part of the data. A company can gain access to a vast user base but loses a slice of the direct relationship and of what it knows about the customer. The transaction fee, as in the model OpenAI describes, is the cost of that presence.

For a small or mid-sized company the takeaway is practical. The point is not to build your own agent overnight, but to avoid disappearing from the shortlist the agent presents to the customer. That is mostly a matter of order in product data and content, so it is achievable without a large budget and without waiting for agentic commerce to become the majority of sales.

Business use cases

Product visibility in agent recommendations

Business problem: a customer asks an AI assistant about a product in the company's category, and the company does not appear because its catalog is unreadable to the machine or incomplete.

AI solution: a structured product feed with clear names, attributes, prices, and availability, plus descriptions that answer real buying questions. OpenAI says Shopify and Etsy catalogs are already integrated with buying in ChatGPT with no extra setup, which shows the direction: data has to be ready for the agent.

Data and processes to implement: a full catalog with attributes, near real-time syncing of prices and stock, structured data on the product page, consistent naming across systems.

Potential effect: presence in the shortlists an agent shows the customer, and retention of sales that used to come through search.

Risk and limitation: some customer data stays with the platform, and the company becomes dependent on its rules and its order of presentation.

Closing the purchase in the agent's channel

Business problem: a product appears in the conversation with an agent, but the customer cannot buy it there and abandons the purchase because they have to move to the store and start over.

AI solution: integration with the agent's buying channel, for example through the Agentic Commerce Protocol from OpenAI and Stripe or Universal Cart and AP2 from Google, while order fulfillment and payment stay on the seller's side.

Data and processes to implement: payment integration, handling of purchase authorization and agentic tokens, clear return and dispute policies, pricing rules consistent with other channels.

Potential effect: a shorter path to purchase and higher conversion from traffic that reaches the company through an AI assistant.

Risk and limitation: the platform fee, less customer data, and unresolved liability rules for a faulty order, which CNBC highlights.

Your own advisory and sales agent on the company site

Business problem: the classic site search and contact form do not lead the customer to a decision, especially for an offer that has to be matched or configured.

AI solution: a company agent that understands intent, asks clarifying questions, and matches the offer, and in B2B can prepare a preliminary quote. Salesforce's Shopper, Buyer, and Merchant agents show how such a division of roles can look in a finished product.

Data and processes to implement: catalog and pricing rules, product knowledge, integration with the CRM and sales system, a mechanism to hand hard cases to a human.

Potential effect: a shorter route from question to offer, service outside working hours, and fewer abandoned conversations.

Risk and limitation: an agent can quote a wrong price or make a wrong promise, so you need quality control over answers, limits on what the agent may commit to, and a clear process owner.

B2B buying and procurement through agents

Business problem: the procurement team compares suppliers by hand and places repeat orders, which eats time and makes it harder to control terms.

AI solution: a procurement agent compares offers, works within set rules, and places orders, which on the supplier side requires an offer that is readable to the agent. Google's AP2 protocol answers the key question here, namely how to prove that a real user authorized the purchase.

Data and processes to implement: current price lists and terms, access through a feed or an application programming interface, approval rules and spend limits, an audit trail for every agent decision.

Potential effect: less time on routine buying and better terms through systematic comparison of offers.

Risk and limitation: the agent commits the company, so without authorization, limits, and audit a costly mistake is easy.

Optimizing content for AI assistant answers

Business problem: the company does not show up in answers from ChatGPT, Perplexity, or Gemini when a customer asks how to solve a problem rather than for a specific product.

AI solution: content that directly answers buying questions, structured data, clear information about the product, delivery, and terms, and question-and-answer sections that make it easier for the model to cite you.

Data and processes to implement: tidy specifications, current availability and price information, consistency between site content and catalog data.

Potential effect: citations and recommendations in assistant answers that send valuable traffic to the company site.

Risk and limitation: the company does not fully control what the model understands from its content, so you have to check regularly how assistants describe the offer and correct the data they rely on.

What companies can do now

Start with a visibility test. Ask ChatGPT, Gemini, and Perplexity a few questions a customer in your category would ask, and check whether the company appears, and if so, whether the description is correct. It is the cheapest way to see how an agent views your offer today.

Clean up product data. A complete catalog with clear names, attributes, price, and availability, kept current, is the foundation of presence in agentic commerce. Without it no integration helps, because the agent has nothing to read.

Name a process owner. Someone has to own how the offer looks in agent channels, the quality of the company agent's answers, and the rules the agent may operate within. Without an owner this area blurs between marketing, sales, and IT.

Start measuring. Break out traffic and sales from AI channels so you know whether agentic commerce actually contributes before you invest in deeper integrations. Decisions about next steps should follow the data, not the pressure that everyone is rolling something out.

Handle security and authorization. If you allow agent purchases, set spend limits, approval rules, and an audit trail. Standards such as AP2 exist precisely to confirm authorization, but the company has to decide what the agent may and may not do.

Risks

Loss of the relationship and customer data. When a purchase goes through the agent's platform, it holds the context and part of the data, and the company sees only the outcome. Access to a large user base can come at the price of thinner knowledge of your own customer.

Dependence on platforms and their fees. Presence in agent channels means operating on the platform's terms, including its presentation rules and charges. That is a cost and a strategic risk to accept deliberately.

Faulty orders and agent commitments. An agent that places orders or promises terms can get it wrong. Liability and dispute rules for agent purchases, as CNBC notes, are not yet resolved, so a company should limit the agent's authority itself.

Shadow AI on the buying side. Employees can use AI assistants for company purchases without procurement knowing, which makes spend and terms harder to control. It is a variant of the familiar unauthorized-AI problem, only with real money.

No owner and no quality control. Without a responsible person and regular checks on what agents say about the offer, divergent prices, stale data, and brand-damaging answers are easy.

Integration cost without return. Agentic commerce sits at the peak of inflated expectations, and only 17 percent of organizations have deployed it according to Gartner. Investing in deep integrations before you see traffic from AI channels may not pay off.

Key takeaways

Commerce is shifting from clicking toward delegating the purchase to an agent, and the choice of product is made first on the machine's side. Salesforce shipped commerce agents as a finished product, Google and the payment networks are building the infrastructure, and the potential scale, per McKinsey, runs into the trillions of dollars. At the same time Gartner is a reminder that this is still an early stage and that liability rules are not settled.

For a company the priority is what can be done now and cheaply: check how AI assistants see the offer, clean up product data, name a process owner, and start measuring the share of AI channels. Deeper integrations and your own agent are the next step, one that should follow the data rather than the rush. The winners are companies that are readable to machines and keep control over price, answer quality, and authorization, not those that hand agents more than they can control.