Agentic Commerce 2026: Building AI Agents That Actually Place Orders (Not Just Chat)
5 Views
Summarize Article
Related Blogs
Summarize Article
On September 16, 2025, Google announced the Agent Payments Protocol (AP2), an open standard built with more than 60 partners including Mastercard, PayPal, and American Express, designed to let an AI agent cryptographically prove a human authorized a specific purchase before money moves, according to Google Cloud’s official announcement. Thirteen days later, Stripe and OpenAI released a competing standard, the Agentic Commerce Protocol, and launched Instant Checkout inside ChatGPT, according to Stripe’s own newsroom post.
Two of the largest technology companies in the world raced to solve the same problem in the same month: a chatbot that recommends a product is not the same thing as an agent that can actually complete the purchase, and until recently, nothing in the payments stack could tell the difference.
That distinction is the entire subject of this article. Most “AI shopping assistant” projects stop at recommendation, because agentic commerce development, the actual engineering of an agent that discovers, carts, authorizes, and pays, requires a protocol layer, a data layer, and a guardrail layer that a conversational interface alone does not provide. This article covers what changed in the past year, what custom AI shopping agents need to work in production, how B2B agentic workflows differ from consumer ones, and what AI-driven checkout optimization actually looks like once the data is in.
A chatbot that describes a product and links to a website is doing recommendation, not commerce. Real agentic commerce development requires three things a plain conversational interface does not have: a way to prove the user actually authorized the specific purchase, a way for the agent and merchant to exchange structured cart and pricing data, and a way to settle payment without exposing raw card credentials to the agent itself. That is exactly the gap AP2 and ACP were built to close, using signed “Mandates” that create a verifiable record of what the user wanted, what the agent selected, and what was actually charged.
A model that can hold a conversation about products is a small fraction of what a working agentic commerce system needs. The rest is unglamorous but non-negotiable: machine-readable product catalogs, real-time inventory and pricing APIs, defined spending limits and category restrictions, and a fallback path for when the agent’s request does not match what the merchant can actually fulfill. Skipping this layer is exactly how a January 2025 research demo of OpenAI’s Operator agent ended up purchasing a dozen eggs from Instacart when the shopping rules it had been given were ambiguous, an incident that became a widely cited example of why authorization protocols exist in the first place.
The shift is not theoretical. Adobe’s holiday 2025 data, based on more than a trillion visits to US retail sites, found AI-referred traffic converting 31% higher than traffic from any other source, nearly double the conversion advantage seen a year earlier, with AI-driven revenue per visit up 254% year over year. By March 2026, that conversion advantage had grown to 42%, a full reversal from March 2025, when AI traffic converted 38% worse than non-AI traffic. Retailers whose catalogs are not machine-readable are the ones being left out of that growth, regardless of how good their human-facing website is.
Deloitte’s 2026 research, based on a survey of more than 1,000 US suppliers and buyers, found 72% of suppliers describing their sales processes as mostly or highly automated, while only 47% of buyers agreed, with buyers six times more likely to call the same processes mostly manual. Suppliers with high digital commerce maturity exceeded their annual sales goals by 110% more than low-maturity competitors and were roughly five times more likely to use AI extensively. The gap between internal automation and what the buyer actually experiences is precisely where B2B agentic workflows either succeed or quietly fail.
Deciding whether your commerce stack is ready for agentic checkout?
WebOsmotic scopes what protocol compliance, catalog readiness, and payment authorization actually require for your platform before any build begins.
Not every vendor offering to “add AI to your storefront” is doing agentic commerce development in the sense this article describes. A few things separate a partner who can actually ship this from one selling a chatbot with extra steps:
A genuine agentic commerce development partner will walk through all five of these before proposing a single line of code, because skipping any of them is exactly how a promising pilot turns into a production incident.
Custom AI shopping agents built without this foundation tend to work fine in a demo and fail unpredictably the moment real inventory, real pricing changes, or a real edge case shows up. This is the exact gap genuine agentic commerce development is supposed to close.
| Factor | Consumer agentic commerce | B2B agentic workflows |
|---|---|---|
| Authorization model | Single-purchase Mandate, often human-present | Standing procurement rules, multi-step approval chains |
| Data requirements | Product catalog, pricing, inventory | Catalog plus contract terms, volume pricing, compliance status |
| Negotiation | Minimal; price is typically fixed | Agent-to-agent quote negotiation within governance guardrails |
| Failure cost | A wrong order, usually returnable | A wrong purchase order, often contractually binding |
| Governance | User-set spending limits | Procurement policy, approval hierarchy, supplier compliance rules |
Deloitte’s research frames the most advanced stage of this shift as procurement agents interacting directly with supplier agents to negotiate pricing, execute transactions, and coordinate fulfillment in real time, with human roles moving from executing every step to setting policy and handling exceptions. That is a materially different engineering problem than a consumer shopping agent, and treating B2B agentic workflows as a smaller version of consumer commerce is a common and costly mistake.
Getting this right is one of the more concrete payoffs of agentic commerce development, since checkout is the exact point where a well-built agent either completes a sale or loses it back to a traditional browser tab.
| Checkout model | Authorization | Merchant integration effort | Conversion impact |
|---|---|---|---|
| Traditional web checkout | Manual, human-driven | N/A (existing baseline) | Baseline |
| Chatbot with outbound link | None; agent recommends, human completes purchase elsewhere | Minimal | Limited, since the agent is not part of the transaction |
| Protocol-compliant agentic checkout (AP2, ACP, UCP) | Cryptographically signed Mandates | Moderate; often “as little as one line of code” for existing Stripe merchants | 31-42% higher conversion from AI referrals, per Adobe |
AI-driven checkout optimization, in practice, is less about tuning a conversion funnel and more about removing the friction of leaving the conversation at all. The data backs that up directly: once a shopper reaches a retail site through an AI referral, they are converting better than shoppers from any other channel, provided the merchant’s systems can actually complete the transaction inside that flow rather than routing the shopper back to a traditional checkout page.
Ready to build agentic commerce into your existing platform?
WebOsmotic delivers agentic commerce development for AP2, ACP, and UCP-compliant checkout, structured catalog data, and B2B workflow automation, for clients who need this built right the first time.
The protocol war is not over, and it may never fully resolve into one winner. What is already settled is the direction: commerce systems that cannot be read and transacted with by an agent are losing traffic, conversion, and, increasingly, entire categories of B2B demand to the ones that can. Agentic commerce development is not a feature to bolt onto an existing storefront. It is infrastructure work, and the companies that treat agentic commerce development that way in 2026 are the ones showing up in the conversion data.
Agentic commerce development is the engineering work required for an AI agent to discover a product, assemble a cart, get verified authorization from a user, and complete payment, rather than simply describing products and linking to a website. It requires implementing at least one commerce protocol (AP2, ACP, or UCP), exposing machine-readable catalog and inventory data, and building an authorization and audit layer that can prove what was purchased and why.
A chatbot answers questions and links out; the human still completes the purchase on a traditional checkout page. Agentic commerce means the agent itself completes the transaction, which requires cryptographic proof of user authorization (via protocols like AP2’s signed Mandates), real-time pricing and inventory data, and a payment settlement path that does not expose raw payment credentials to the agent. This is the specific line agentic commerce development work has to cross that a chatbot integration never needs to.
Standing procurement rules and multi-step approval chains rather than a single authorization, contract-aware pricing rather than fixed catalog prices, and often agent-to-agent negotiation within governance guardrails. Deloitte’s research found 72% of suppliers consider their processes automated versus only 47% of buyers, which shows the gap is less about technology availability and more about how the workflow is actually built and governed.
Not necessarily all three, but supporting none of them means being effectively invisible to agent-mediated discovery and checkout. Most merchants start with whichever protocol aligns with their existing payment processor, Stripe merchants can often enable ACP with minimal code changes, and expand coverage as adoption data shows where their specific customers are transacting.
Mostly, but the mechanism is different from traditional conversion rate optimization. Adobe’s data shows AI-referred traffic already converting 31 to 42% better than other channels once shoppers reach a retail site. The optimization work is less about persuading the shopper and more about making sure the technical path from agent recommendation to completed purchase does not force the shopper back into a separate, friction-heavy checkout flow.
