Uncategorized

Agentic Commerce Explained: The Future of AI Shopping

Photo of Daniel Carter45 min read

Introduction: What Happens When Your Next Customer Is an AI Agent?

A shopper in Delhi opens an AI assistant and types: “Find me a reliable laptop under ₹70,000 with good battery life, compare the available options, and help me purchase the best one.”

A few years ago that sentence would have launched a dozen browser tabs. Today, AI-powered shopping assistants can interpret the request, pull candidates from product feeds and search indexes, set battery claims against independent reviews, notice that one model has a 7-day return window while another has 30, and return a short list with reasons. The shopper picks one. What happens next depends on the assistant, the country, and the merchant, and that is the part most articles skip.

In some US flows, the assistant can pass a prepared cart into a checkout the merchant controls, and the customer approves payment in a few taps. In others, it simply hands the shopper to the merchant’s own site or app. As of October 31, 2026, assistants widely support discovery and comparison, while agent-initiated checkout is limited by region, merchant eligibility, and payment infrastructure. For Indian merchants, the practical opportunity today is being discoverable and trustworthy to agents, while building toward checkout and payments as those rails open.

This is agentic commerce: software agents acting on a customer’s behalf across discovery, comparison, and, with explicit authorization, purchase. For readers of ValuFlash, whether you run a store, build SaaS, or work in payments, the opportunity and the challenge are the same thing. Your catalog, APIs, policies, and payment flows must be readable by software that evaluates you before a human ever sees your homepage.

This guide separates announced concepts from generally available products, explains how agentic AI in e-commerce works, covers the main protocols, and walks through a buildable implementation for a fictional Indian electronics retailer.

What Is Agentic Commerce?

The beginner-friendly definition

Agentic commerce is online shopping in which an AI agent does part of the work a customer would normally do: understanding what they want, finding products, comparing them, and, only when the customer authorizes it, helping complete the purchase.

The technical definition

Technically, it is an architecture in which a software agent with delegated, scoped authority calls merchant and payment systems through machine-readable interfaces (APIs, product feeds, tool servers) to execute steps of a purchase on a customer’s behalf, with identity, authorization, and audit controls around each step. It is a systems problem as much as an AI problem. Models reason, but catalogs, checkout sessions, payment tokens, and order records still do the work.

Related, but not interchangeable

ModelWho does the workWho authorizes the purchaseExampleMaturity (Oct 2026)
Traditional e-commerceCustomer browses, filters, checks outCustomer at checkoutVisiting a store websiteMature
Conversational commerceCustomer chats with a brand’s own bot or messaging flowCustomerBrand chatbot, messaging-app orderingMature, brand-controlled
AI-assisted shoppingAI researches and compares; customer buys elsewhereCustomer on the merchant’s siteAI answer listing products with linksWidely available
Agentic commerceAgent discovers, compares, builds a cart, may start checkout through supported interfacesCustomer, explicitlyCart handed to a supported checkoutLimited; varies by platform, region, merchant
Fully autonomous purchasingAgent buys inside pre-set rules without per-purchase approvalCustomer, through standing rulesRecurring bill payment under a capEarly; pilots and announced frameworks

These are related but not interchangeable. A chatbot that answers questions about one brand’s catalog is conversational commerce, not agentic commerce, because it does not act across merchants or hold delegated authority. An assistant that recommends products and links out is AI-assisted shopping. And an assistant that can complete a purchase without asking each time is a much bigger step, requiring payment authorization, liability, and identity infrastructure that is still being standardized. Not every AI assistant can independently complete purchases, and treating them as if they can leads to bad planning.

How Agentic Commerce Actually Works

The conceptual workflow has ten steps. Which steps are automated depends on the platform and merchant.

  1. Customer expresses purchasing intent. “Wireless headphones under ₹5,000, strong battery, easy returns.”
  2. The agent interprets requirements. It separates hard constraints (budget, category) from soft preferences (battery life) and notes what is ambiguous.
  3. The agent discovers merchants and products. It queries catalogs, feeds, search indexes, or merchant tool servers.
  4. Product data is retrieved. Titles, identifiers, attributes, price, images, policies.
  5. Products are compared. Against the customer’s criteria, with reasons.
  6. Availability, shipping, and price are checked. Ideally against live merchant systems, not a cached copy.
  7. Customer authorization is obtained. An explicit approval step, or a narrowly scoped prior authorization where supported.
  8. Checkout is initiated through supported infrastructure. A checkout session on the merchant’s systems, or a handoff to the merchant’s own checkout.
  9. Payment is processed securely. Through the merchant’s payment provider using scoped credentials, not raw card details handed to the agent.
  10. Order confirmation and fulfillment information are returned. Order ID, status, tracking, and return instructions.

Who does what

Step areaAI agent / platformMerchant systemsPayment provider
Understanding intentInterprets, asks clarifying questionsn/an/a
Discovery and dataQueries and ranksPublishes accurate catalog and policiesn/a
Price, stock, shippingRequests and displaysSource of truth; calculates tax and shippingn/a
AuthorizationPresents a summary and requests approvalValidates the order it will fulfilIssues or validates scoped credentials
Checkout and paymentPasses the customer’s approved intentOwns checkout state and the orderAuthorizes, captures, settles
Post-purchaseRelays statusFulfils; handles returns and disputesHandles refunds and chargebacks

The key principle: the agent proposes and relays, the merchant remains the source of truth for price, stock, and the order, and the payment provider controls money movement.

Traditional E-commerce vs Agentic Commerce

DimensionTraditional e-commerceAgent-assisted commerce
Product discoverySite search, category browsing, adsNatural-language requests resolved against feeds, catalogs, and indexes
Search behaviorShort keyword queriesLong, constraint-rich requests (“under ₹70,000, 8+ hours battery”)
Customer interfaceMerchant’s website or appAssistant interface; merchant site or app may still handle checkout
PersonalizationOn-site behavior and CRMAssistant’s knowledge of stated preferences; merchant sees less context
Product comparisonCustomer opens tabs, reads specsAgent produces side-by-side comparisons from structured data and reviews
CheckoutMerchant-designed flowVaries: link-out, cart transfer, or in-surface checkout where supported
Payment authorizationCustomer enters or selects a payment methodCustomer approves; scoped tokens or delegated credentials where supported
Merchant integrationSite, ads, marketplacesFeeds, APIs, and protocol endpoints per platform
SEO and GEORankings, snippetsData accuracy, structured attributes, answer-engine visibility
SecurityBot defense, payment fraud, account takeoverAdds prompt injection, agent identity, delegated-authority abuse
Customer relationshipDirect, merchant-ownedMediated by the assistant; merchant must earn it at checkout and post-purchase

Agent-assisted commerce is more likely to complement existing sites and apps than replace them. Several major platforms deliberately send shoppers back to merchant checkout, as the sections below show.

Why Agentic Commerce Matters for Businesses

Product discovery is changing. Shoppers increasingly start with a conversational request instead of a keyword. Merchants whose data cannot answer constraint-based questions risk being skipped.

New acquisition channels are forming, with uncertain economics. Assistants and platform surfaces can send qualified traffic. No verified industry benchmarks exist yet, so treat any claimed uplift or lower acquisition cost as unproven until you measure it on your own store.

Product-data quality becomes a competitive asset. Trade press reported in March 2026 that Walmart said purchases inside ChatGPT converted at about one-third the rate of clicks to its own site, with product-data inaccuracy cited as an issue. We could not retrieve a primary Walmart source, so read this as one reported data point, not a benchmark. The lesson holds regardless: stale prices and incomplete attributes break agent flows.

Customer experience shifts. The first impression may be an assistant’s summary of your product, not your homepage. Your policies, reviews, and specs are what get summarized.

Merchant competition intensifies. When an agent compares ten merchants on price, delivery date, and return window in seconds, weak policies and opaque pricing become visible.

Platform dependency is a real risk. Protocols, eligibility rules, and surfaces change quickly. OpenAI moved away from standalone in-chat checkout within about six months of launching it (details below). Build on your own data and infrastructure so you can adapt.

Pricing transparency increases. Agents surface total cost, shipping, and policies up front. Hidden fees hurt more.

Operations get harder. You will need inventory accuracy, reconciliation for agent-originated orders, and clear handling of returns and disputes. Solid e-commerce automation (stock sync, order tagging, return workflows) becomes a prerequisite, not a luxury.

Agentic Commerce Architecture Explained

The diagram below is a conceptual reference architecture for AI commerce infrastructure, not any single vendor’s design. It shows four trust zones and the point where customer authorization must occur. This kind of design sits inside the broader shift toward agent infrastructure that we track in ValuFlash’s AI coverage. #mermaid-rv1-r1-2e9bab9d7b86bdaa { font-family: system-ui, “Segoe UI”, Roboto, Helvetica, Arial, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, “Kohinoor Devanagari”, “Kohinoor Bangla”, “Kohinoor Telugu”, “Tamil Sangam MN”, “Kohinoor Gujarati”, “Malayalam Sangam MN”, “Nirmala UI”, “Noto Sans Devanagari UI”, “Noto Sans Devanagari”, “Noto Sans Bengali UI”, “Noto Sans Bengali”, “Noto Sans Telugu UI”, “Noto Sans Telugu”, “Noto Sans Tamil UI”, “Noto Sans Tamil”, “Noto Sans Gujarati UI”, “Noto Sans Gujarati”, “Noto Sans Kannada UI”, “Noto Sans Kannada”, “Noto Sans Malayalam UI”, “Noto Sans Malayalam”, Thonburi, “Leelawadee UI”, “Noto Sans Thai UI”, “Noto Sans Thai”, Kefa, Ebrima, “Noto Sans Ethiopic”, “Abyssinica SIL”, sans-serif, ui-sans-serif, system-ui, -apple-system, “Segoe UI”, Roboto, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, sans-serif; font-size: 16px; fill: rgb(229, 229, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-animation-slow { stroke-dashoffset: 900; animation: 50s linear 0s infinite normal none running dash; stroke-linecap: round; stroke-dasharray: 9, 5 !important; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-animation-fast { stroke-dashoffset: 900; animation: 20s linear 0s infinite normal none running dash; stroke-linecap: round; stroke-dasharray: 9, 5 !important; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .error-icon { fill: rgb(204, 120, 92); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .error-text { fill: rgb(51, 135, 163); stroke: rgb(51, 135, 163); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-thickness-normal { stroke-width: 1px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-thickness-thick { stroke-width: 3.5px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-pattern-solid { stroke-dasharray: 0; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-thickness-invisible { stroke-width: 0; fill: none; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-pattern-dashed { stroke-dasharray: 3; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edge-pattern-dotted { stroke-dasharray: 2; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .marker { fill: rgb(161, 161, 161); stroke: rgb(161, 161, 161); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .marker.cross { stroke: rgb(161, 161, 161); } #mermaid-rv1-r1-2e9bab9d7b86bdaa svg { font-family: system-ui, “Segoe UI”, Roboto, Helvetica, Arial, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, “Kohinoor Devanagari”, “Kohinoor Bangla”, “Kohinoor Telugu”, “Tamil Sangam MN”, “Kohinoor Gujarati”, “Malayalam Sangam MN”, “Nirmala UI”, “Noto Sans Devanagari UI”, “Noto Sans Devanagari”, “Noto Sans Bengali UI”, “Noto Sans Bengali”, “Noto Sans Telugu UI”, “Noto Sans Telugu”, “Noto Sans Tamil UI”, “Noto Sans Tamil”, “Noto Sans Gujarati UI”, “Noto Sans Gujarati”, “Noto Sans Kannada UI”, “Noto Sans Kannada”, “Noto Sans Malayalam UI”, “Noto Sans Malayalam”, Thonburi, “Leelawadee UI”, “Noto Sans Thai UI”, “Noto Sans Thai”, Kefa, Ebrima, “Noto Sans Ethiopic”, “Abyssinica SIL”, sans-serif, ui-sans-serif, system-ui, -apple-system, “Segoe UI”, Roboto, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, sans-serif; font-size: 16px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa p { margin: 0px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label { font-family: system-ui, “Segoe UI”, Roboto, Helvetica, Arial, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, “Kohinoor Devanagari”, “Kohinoor Bangla”, “Kohinoor Telugu”, “Tamil Sangam MN”, “Kohinoor Gujarati”, “Malayalam Sangam MN”, “Nirmala UI”, “Noto Sans Devanagari UI”, “Noto Sans Devanagari”, “Noto Sans Bengali UI”, “Noto Sans Bengali”, “Noto Sans Telugu UI”, “Noto Sans Telugu”, “Noto Sans Tamil UI”, “Noto Sans Tamil”, “Noto Sans Gujarati UI”, “Noto Sans Gujarati”, “Noto Sans Kannada UI”, “Noto Sans Kannada”, “Noto Sans Malayalam UI”, “Noto Sans Malayalam”, Thonburi, “Leelawadee UI”, “Noto Sans Thai UI”, “Noto Sans Thai”, Kefa, Ebrima, “Noto Sans Ethiopic”, “Abyssinica SIL”, sans-serif, ui-sans-serif, system-ui, -apple-system, “Segoe UI”, Roboto, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, sans-serif; color: rgb(229, 229, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .cluster-label text { fill: rgb(51, 135, 163); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .cluster-label span { color: rgb(51, 135, 163); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .cluster-label span p { background-color: transparent; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label text, #mermaid-rv1-r1-2e9bab9d7b86bdaa span { fill: rgb(229, 229, 229); color: rgb(229, 229, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .node rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .node circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .node ellipse, #mermaid-rv1-r1-2e9bab9d7b86bdaa .node polygon, #mermaid-rv1-r1-2e9bab9d7b86bdaa .node path { fill: transparent; stroke: rgb(161, 161, 161); stroke-width: 1px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .rough-node .label text, #mermaid-rv1-r1-2e9bab9d7b86bdaa .node .label text, #mermaid-rv1-r1-2e9bab9d7b86bdaa .image-shape .label, #mermaid-rv1-r1-2e9bab9d7b86bdaa .icon-shape .label { text-anchor: middle; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .node .katex path { fill: rgb(0, 0, 0); stroke: rgb(0, 0, 0); stroke-width: 1px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .rough-node .label, #mermaid-rv1-r1-2e9bab9d7b86bdaa .node .label, #mermaid-rv1-r1-2e9bab9d7b86bdaa .image-shape .label, #mermaid-rv1-r1-2e9bab9d7b86bdaa .icon-shape .label { text-align: center; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .node.clickable { cursor: pointer; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .root .anchor path { stroke-width: 0; stroke: rgb(161, 161, 161); fill: rgb(161, 161, 161) !important; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .arrowheadPath { fill: rgb(11, 11, 11); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edgePaths .path { stroke: rgb(161, 161, 161); stroke-width: 1px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .flowchart-link { stroke: rgb(161, 161, 161); fill: none; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edgeLabel { background-color: transparent; text-align: center; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edgeLabel p { background-color: transparent; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .edgeLabel rect { opacity: 0.5; background-color: transparent; fill: transparent; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .labelBkg { background-color: rgba(0, 0, 0, 0.5); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .cluster rect { fill: rgb(32, 32, 31); stroke: rgb(161, 161, 161); stroke-width: 1px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .cluster text { fill: rgb(51, 135, 163); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .cluster span { color: rgb(51, 135, 163); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .node .collapsed-indicator { fill: rgb(161, 161, 161); stroke: none; opacity: 0.6; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .node .collapsed-separator { stroke: rgb(161, 161, 161); stroke-width: 0.75px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa div.mermaidTooltip { position: absolute; text-align: center; max-width: 200px; padding: 2px; font-family: system-ui, “Segoe UI”, Roboto, Helvetica, Arial, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, “Kohinoor Devanagari”, “Kohinoor Bangla”, “Kohinoor Telugu”, “Tamil Sangam MN”, “Kohinoor Gujarati”, “Malayalam Sangam MN”, “Nirmala UI”, “Noto Sans Devanagari UI”, “Noto Sans Devanagari”, “Noto Sans Bengali UI”, “Noto Sans Bengali”, “Noto Sans Telugu UI”, “Noto Sans Telugu”, “Noto Sans Tamil UI”, “Noto Sans Tamil”, “Noto Sans Gujarati UI”, “Noto Sans Gujarati”, “Noto Sans Kannada UI”, “Noto Sans Kannada”, “Noto Sans Malayalam UI”, “Noto Sans Malayalam”, Thonburi, “Leelawadee UI”, “Noto Sans Thai UI”, “Noto Sans Thai”, Kefa, Ebrima, “Noto Sans Ethiopic”, “Abyssinica SIL”, sans-serif, ui-sans-serif, system-ui, -apple-system, “Segoe UI”, Roboto, “PingFang SC”, “PingFang TC”, “Hiragino Sans”, “Apple SD Gothic Neo”, sans-serif; font-size: 12px; background: rgb(204, 120, 92); border: 1px solid rgb(138, 115, 107); border-radius: 2px; pointer-events: none; z-index: 100; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .flowchartTitleText { text-anchor: middle; font-size: 18px; fill: rgb(229, 229, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa rect.text { fill: none; stroke-width: 0; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .icon-shape, #mermaid-rv1-r1-2e9bab9d7b86bdaa .image-shape { background-color: transparent; text-align: center; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .icon-shape p, #mermaid-rv1-r1-2e9bab9d7b86bdaa .image-shape p { background-color: transparent; padding: 2px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .icon-shape .label rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .image-shape .label rect { opacity: 0.5; background-color: transparent; fill: transparent; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label-icon { display: inline-block; height: 1em; overflow: visible; vertical-align: -0.125em; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .node .label-icon path { fill: currentcolor; stroke: revert; stroke-width: revert; } #mermaid-rv1-r1-2e9bab9d7b86bdaa .node .neo-node { stroke: rgb(161, 161, 161); } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].node rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].cluster rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].node polygon { stroke: url(“#mermaid-rv1-r1-2e9bab9d7b86bdaa-gradient”); filter: drop-shadow(rgb(185, 185, 185) 1px 2px 2px); } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].swimlane.cluster rect { filter: none; } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].node path { stroke: url(“#mermaid-rv1-r1-2e9bab9d7b86bdaa-gradient”); stroke-width: 1px; } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].node .outer-path { filter: drop-shadow(rgb(185, 185, 185) 1px 2px 2px); } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].node .neo-line path { stroke: rgb(161, 161, 161); filter: none; } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].node circle { stroke: url(“#mermaid-rv1-r1-2e9bab9d7b86bdaa-gradient”); filter: drop-shadow(rgb(185, 185, 185) 1px 2px 2px); } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].node circle .state-start { fill: rgb(0, 0, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].icon-shape .icon { fill: url(“#mermaid-rv1-r1-2e9bab9d7b86bdaa-gradient”); filter: drop-shadow(rgb(185, 185, 185) 1px 2px 2px); } #mermaid-rv1-r1-2e9bab9d7b86bdaa [data-look=”neo”].icon-shape .icon-neo path { stroke: url(“#mermaid-rv1-r1-2e9bab9d7b86bdaa-gradient”); filter: drop-shadow(rgb(185, 185, 185) 1px 2px 2px); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit0, #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight0 { fill: rgb(151, 149, 141); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit0, #mermaid-rv1-r1-2e9bab9d7b86bdaa .arrow0, #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight0 { stroke: rgb(151, 149, 141); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label0 { fill: rgba(255, 255, 255, 0.05); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label0 { fill: rgb(195, 194, 183); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight1 { fill: rgb(57, 135, 229); stroke: rgb(57, 135, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label1 { fill: rgb(3, 32, 66); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label1 { fill: rgb(109, 167, 236); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight2 { fill: rgb(217, 89, 38); stroke: rgb(217, 89, 38); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label2 { fill: rgb(55, 20, 7); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label2 { fill: rgb(236, 131, 90); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight3 { fill: rgb(25, 158, 112); stroke: rgb(25, 158, 112); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label3 { fill: rgb(2, 39, 32); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label3 { fill: rgb(59, 189, 140); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight4 { fill: rgb(201, 133, 0); stroke: rgb(201, 133, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label4 { fill: rgb(49, 26, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label4 { fill: rgb(219, 147, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight5 { fill: rgb(213, 81, 129); stroke: rgb(213, 81, 129); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label5 { fill: rgb(57, 15, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label5 { fill: rgb(232, 123, 164); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight6 { fill: rgb(0, 131, 0); stroke: rgb(0, 131, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label6 { fill: rgb(17, 38, 15); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label6 { fill: rgb(85, 191, 80); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight7 { fill: rgb(144, 133, 233); stroke: rgb(144, 133, 233); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .label7 { fill: rgb(29, 22, 73); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .branch-label7 { fill: rgb(160, 150, 235); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-merge, #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-reverse, #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-highlight-inner { fill: rgb(45, 45, 45); stroke: rgb(45, 45, 45); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-cherry-pick[fill], #mermaid-rv1-r1-2e9bab9d7b86bdaa .commit-cherry-pick[stroke] { fill: rgb(45, 45, 45); stroke: rgb(45, 45, 45); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section–1 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section–1 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section–1 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section–1 polygon { fill: rgb(3, 32, 66); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section–1 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section–1 .label-container { stroke: rgb(57, 135, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section–1 text { fill: rgb(109, 167, 236); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-0 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-0 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-0 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-0 polygon { fill: rgb(55, 20, 7); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-0 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-0 .label-container { stroke: rgb(217, 89, 38); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-0 text { fill: rgb(236, 131, 90); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-1 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-1 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-1 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-1 polygon { fill: rgb(2, 39, 32); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-1 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-1 .label-container { stroke: rgb(25, 158, 112); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-1 text { fill: rgb(59, 189, 140); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-2 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-2 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-2 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-2 polygon { fill: rgb(49, 26, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-2 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-2 .label-container { stroke: rgb(201, 133, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-2 text { fill: rgb(219, 147, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-3 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-3 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-3 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-3 polygon { fill: rgb(57, 15, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-3 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-3 .label-container { stroke: rgb(213, 81, 129); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-3 text { fill: rgb(232, 123, 164); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-4 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-4 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-4 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-4 polygon { fill: rgb(17, 38, 15); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-4 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-4 .label-container { stroke: rgb(0, 131, 0); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-4 text { fill: rgb(85, 191, 80); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-5 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-5 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-5 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-5 polygon { fill: rgb(29, 22, 73); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-5 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-5 .label-container { stroke: rgb(144, 133, 233); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-5 text { fill: rgb(160, 150, 235); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-6 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-6 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-6 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-6 polygon { fill: rgb(60, 14, 14); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-6 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-6 .label-container { stroke: rgb(230, 103, 103); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-6 text { fill: rgb(236, 126, 126); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-7 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-7 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-7 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-7 polygon { fill: rgb(32, 32, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-7 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-7 .label-container { stroke: rgb(95, 94, 90); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-7 text { fill: rgb(195, 194, 183); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-8 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-8 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-8 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-8 polygon { fill: rgb(32, 32, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-8 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-8 .label-container { stroke: rgb(95, 94, 90); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-8 text { fill: rgb(195, 194, 183); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-9 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-9 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-9 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-9 polygon { fill: rgb(32, 32, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-9 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-9 .label-container { stroke: rgb(95, 94, 90); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-9 text { fill: rgb(195, 194, 183); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-10 rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-10 path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-10 circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-10 polygon { fill: rgb(32, 32, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-10 .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-10 .label-container { stroke: rgb(95, 94, 90); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-10 text { fill: rgb(195, 194, 183); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-root rect, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-root path, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-root circle, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-root polygon { fill: rgb(32, 32, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-root .node-bkg, #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-root .label-container { stroke: rgb(151, 149, 141); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .section-root text { fill: rgb(195, 194, 183); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .items .node rect { fill: rgb(32, 32, 31); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .critText0, #mermaid-rv1-r1-2e9bab9d7b86bdaa .critText1, #mermaid-rv1-r1-2e9bab9d7b86bdaa .critText2, #mermaid-rv1-r1-2e9bab9d7b86bdaa .critText3 { fill: rgb(255, 255, 255); } #mermaid-rv1-r1-2e9bab9d7b86bdaa text.journey-section { fill: rgb(229, 229, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .titleText, #mermaid-rv1-r1-2e9bab9d7b86bdaa .sectionTitle0, #mermaid-rv1-r1-2e9bab9d7b86bdaa .sectionTitle1, #mermaid-rv1-r1-2e9bab9d7b86bdaa .sectionTitle2, #mermaid-rv1-r1-2e9bab9d7b86bdaa .sectionTitle3 { fill: rgb(229, 229, 229); } #mermaid-rv1-r1-2e9bab9d7b86bdaa rect.treemapSection { fill-opacity: 0.3; } #mermaid-rv1-r1-2e9bab9d7b86bdaa text[fill=”#444444″] { fill: rgb(229, 229, 229); stroke: rgb(26, 26, 25); stroke-width: 3px; paint-order: stroke; } #mermaid-rv1-r1-2e9bab9d7b86bdaa line[stroke=”#444444″], #mermaid-rv1-r1-2e9bab9d7b86bdaa path[stroke=”#444444″][marker-end], #mermaid-rv1-r1-2e9bab9d7b86bdaa path[stroke=”#444444″][marker-start], #mermaid-rv1-r1-2e9bab9d7b86bdaa rect[stroke=”#444444″] { stroke: rgb(161, 161, 161); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .block .node .cluster, #mermaid-rv1-r1-2e9bab9d7b86bdaa .statediagram-state rect.divider { stroke: rgb(161, 161, 161); } #mermaid-rv1-r1-2e9bab9d7b86bdaa .mindmap-node .label { transform: translate(-50%, -50%); transform-box: fill-box; } #mermaid-rv1-r1-2e9bab9d7b86bdaa :root { –mermaid-font-family: system-ui,”Segoe UI”,Roboto,Helvetica,Arial,”PingFang SC”,”PingFang TC”,”Hiragino Sans”,”Apple SD Gothic Neo”,”Kohinoor Devanagari”,”Kohinoor Bangla”,”Kohinoor Telugu”,”Tamil Sangam MN”,”Kohinoor Gujarati”,”Malayalam Sangam MN”,”Nirmala UI”,”Noto Sans Devanagari UI”,”Noto Sans Devanagari”,”Noto Sans Bengali UI”,”Noto Sans Bengali”,”Noto Sans Telugu UI”,”Noto Sans Telugu”,”Noto Sans Tamil UI”,”Noto Sans Tamil”,”Noto Sans Gujarati UI”,”Noto Sans Gujarati”,”Noto Sans Kannada UI”,”Noto Sans Kannada”,”Noto Sans Malayalam UI”,”Noto Sans Malayalam”,Thonburi,”Leelawadee UI”,”Noto Sans Thai UI”,”Noto Sans Thai”,Kefa,Ebrima,”Noto Sans Ethiopic”,”Abyssinica SIL”,sans-serif,ui-sans-serif,system-ui,-apple-system,”Segoe UI”,Roboto,”PingFang SC”,”PingFang TC”,”Hiragino Sans”,”Apple SD Gothic Neo”,sans-serif; }Zone 4 – Payment and identityZone 3 – Merchant systemsZone 2 – Agent platformZone 1 – Customer controlledcheckout requestEXPLICIT APPROVAL GATEpayment requestresultconfirmationCustomer – intent,preferences, spend limitsChat or app interfaceAI agent and orchestrationMerchant and productdiscoveryProduct catalog APIInventory and pricingservicesShipping and tax calculationCheckout serviceOrder managementFulfilmentPayment authorization -scoped token or mandatePayment processorLogging and monitoringacross all zones

Reading the architecture

  • Customer interface and intent. Preferences, budget, and limits live here. This is the only place a customer can truly approve a purchase.
  • Agent and orchestration. Plans steps, calls tools, and explains recommendations. It should hold the least privilege necessary and never be the source of truth for price or stock.
  • Merchant discovery. Feeds, indexes, or protocol-level discovery documents let the agent find merchants and read their capabilities.
  • Catalog, inventory, pricing, shipping, and tax. Merchant-owned services. The agent must re-verify them immediately before checkout.
  • Checkout service. Creates a session with line items, totals, and status. It enforces business rules and rejects invalid orders.
  • Payment authorization and processor. Credentials are scoped (single merchant, capped amount, expiry) and processed by the merchant’s payment provider.
  • Order management and fulfillment. The same systems as today, plus a source tag so agent-originated orders can be tracked.
  • Logging and monitoring. Records who requested what, which price was shown, and which approval was given.

Trust boundaries. Every arrow crossing zones needs authentication, input validation, and logging. Treat content from Zone 2 and any third-party product text as untrusted input. Authorization must occur at the approval gate, with the exact items, total, and merchant shown to the customer, before any credential is released.

Agentic Commerce Protocols and Emerging Standards

Several protocols exist, built by different organizations at different layers. All status notes below were checked on October 31, 2026 and will age quickly. Check each primary source before you build.

Agentic Commerce Protocol (ACP)

  • Purpose. An open standard for connecting buyers, their AI agents, and businesses to complete purchases. It defines a checkout specification and a delegated payment specification, and OpenAI has extended it to product discovery through merchant-supplied feeds.
  • Organizations. Maintained by OpenAI and Stripe under an Apache 2.0 license; first released September 29, 2025. PayPal announced adoption on October 28, 2025.
  • Capabilities. The merchant implements endpoints that let the agent platform create and update a checkout session. Per OpenAI’s documentation, the checkout session is rendered in OpenAI’s interface, but checkout state and payment processing remain on the merchant’s systems, and OpenAI is not the merchant of record. For payments, OpenAI prepares a one-time delegated payment request with a maximum chargeable amount and an expiry, which a payment provider turns into a token for the merchant. Stripe’s Shared Payment Token is the first compatible implementation.
  • Current documented status. The GitHub specification is marked beta, with a dated snapshot of 2026-04-17. OpenAI’s documentation says Instant Checkout in ChatGPT is available to approved partners. At ShopTalk on March 24, 2026, OpenAI said it was letting merchants use their own checkout while it focused on product discovery, and its merchant page now says it is moving away from a standalone Instant Checkout. Trade press estimates that only roughly a dozen to about 30 Shopify merchants ever went live on it. We could not retrieve OpenAI’s own product-discovery announcement directly, so those details come from secondary reporting.
  • Merchant implications. Feeds matter most today. A checkout integration only makes sense if you are an approved partner or your platform handles it.
  • Limitations. Beta status, a narrow checkout path, and fast-moving platform strategy.

Universal Commerce Protocol (UCP)

  • Purpose. An open standard for the full commerce lifecycle between agents, merchants, payment providers, and credential providers.
  • Organizations. Announced by Google on January 11, 2026, co-developed with Shopify, Etsy, Wayfair, Target, and Walmart; Apache 2.0, published at ucp.dev.
  • Capabilities. Modular capabilities (checkout, orders, catalog, and cart) with extensions, negotiated per transaction. Businesses publish a profile at a well-known path, and capabilities can be offered over REST, MCP, or A2A. Google states UCP is compatible with AP2, A2A, and MCP.
  • Current documented status. Google’s Merchant Center help page says UCP-powered checkout applies only to products with eligibility in the United States, Canada, and Australia, for select participating merchants, via an interest form and a technical build. Merchants must also set the native_commerce(checkout_eligibility) product attribute to show a Buy button, and the merchant remains seller of record. Google launched Universal Cart in the US across Search and the Gemini app on October 15, 2026, and said UCP checkout will expand to Canada and Australia in the coming months and later the UK. India is not listed for UCP checkout. On Shopify’s side, UCP-compliant MCP servers for catalog, cart, checkout, and order tracking are documented, and a Universal Cart API is offered through an early-access waitlist.
  • Merchant implications. Shopify merchants get much of this through the platform. Others need a protocol build, a Merchant Center account in good standing, and approval.
  • Limitations. Select-merchant access, limited countries, and Google Wallet card credentials as the current payment path.

Agent Payments Protocol (AP2)

  • Purpose. An authorization layer: cryptographically signed mandates that prove what a user authorized an agent to buy, creating an audit trail for merchants, processors, and customers.
  • Organizations. Created by Google (announced September 2025). On April 28, 2026 Google donated AP2 to the FIDO Alliance and published AP2 v0.2, which adds “Human Not Present” autonomous payments. Mastercard’s Verifiable Intent framework, co-developed with Google, was donated as well.
  • Current documented status. A published specification with reference code. On May 19, 2026, Google said it would begin bringing AP2 to Google products “in the coming months,” starting with Gemini Spark. That is an announced rollout, not a general availability statement. Earlier versions described intent, cart, and payment mandates; later reporting says v0.2 restructures around checkout and payment mandates, so read the current spec.
  • Merchant implications. Mostly handled by your payment provider and agent platform, but your checkout must be able to carry and validate authorization evidence.
  • Limitations. A young specification, not yet proven at card-network volume, and it does not solve agent identity on its own.

Network and national rails

Visa (Intelligent Commerce, Trusted Agent Protocol) and Mastercard (Agent Pay) work on agent identity and tokenized credentials. Trade reports describe live pilots, including Visa transactions at European merchants from July 2, 2026 and a Visa pilot where a human still approves each invoice. We did not retrieve Visa’s or Mastercard’s own pages directly, so treat these as reported. In India, NPCI announced a Unified Agent Protocol at Global Fintech Fest in September 2026 as “coming,” building on UPI Circle and Reserve Pay. Third-party checks as of early October 2026 found no public specification or go-live date.

Comparison

Protocol or programMain backersLayerDocumented status (checked Oct 31, 2026)Main limitation
ACPOpenAI, StripeCheckout, delegated payment, product feedsSpec in beta; Instant Checkout for approved partners; OpenAI emphasizes discoveryNarrow checkout path; strategy has shifted
UCPGoogle, Shopify, retail co-developersCatalog, cart, checkout, ordersOpen spec; Google checkout for select merchants in US, Canada, Australia; Universal Cart live in USNot available for Google checkout in India
AP2Google, now FIDO AlliancePayment authorization (mandates)v0.2 published; Google rollout to its products announcedEarly; needs surrounding commerce protocol
Visa Intelligent Commerce / TAPVisaAgent identity, tokensDeveloper release Oct 2025; pilots reportedReaches merchants via processors
Mastercard Agent PayMastercardAgentic tokens, agent verificationPilots and live transactions reported in some marketsReaches merchants via issuers and partners
NPCI Unified Agent ProtocolNPCI (India)UPI delegated paymentsAnnounced as coming (Sept 2026); no public spec foundSpecification and liability model unpublished

These are not interchangeable. Supporting ACP does not make you UCP-compatible, and AP2 does not replace a checkout protocol. Some payment providers say they can accept transactions initiated through several of them, and platforms like Shopify abstract the differences. Do not assume universal interoperability, and do not claim any merchant supports a protocol without evidence from that merchant.

How AI Agents Discover Products

AI product discovery works through a mix of search indexes, merchant-submitted feeds, platform catalogs, and, increasingly, protocol-level catalog endpoints. OpenAI’s documentation describes a secure, regularly refreshed feed in CSV or JSON containing identifiers, descriptions, pricing, inventory, media, and fulfillment options, validated with a sample feed and then updated with daily snapshots. Google’s Merchant Center remains the hub for Google surfaces, and Shopify syndicates eligible products through Shopify Catalog.

What agents rely on:

  • Product feeds and APIs. The machine-readable source of truth.
  • Structured data. Schema.org Product and Offer markup helps crawlers and retrieval systems parse your pages.
  • Product identifiers. GTINs, MPNs, and SKUs let systems match your listing to the same item elsewhere.
  • Descriptions and category information. Clear titles and correct categories decide which queries you match.
  • Attributes. Battery hours, screen size, material, compatibility. These answer constraint-based requests.
  • Inventory and pricing. Stale values create failed checkouts and lost trust.
  • Merchant policies. Shipping times, return windows, warranty, and support contacts feed directly into recommendations.

Why incomplete data breaks automated shopping. A human forgives a missing spec by calling support. An agent either skips the product, guesses, or presents a wrong claim. Inconsistent variants (the “128 GB” listing that is actually 64 GB), mismatched prices between feed and checkout, and missing return policies are exactly the failures that drive the conversion gaps trade press has described.

SEO vs GEO vs Agentic Commerce Optimization

Technical SEOGEO (generative engine optimization)Agentic commerce optimization
GoalRank pages in search resultsBe cited accurately in AI answersBe selected, compared, and transacted with by agents
Primary inputsCrawlable pages, links, speed, markupClear, factual, attributable contentFeeds, APIs, identifiers, live inventory and price, policies
Failure modeLow rankingMissing or misquoted in answersSkipped, mismatched, or failed at checkout

These overlap, but the third adds transactional reliability. No platform has published a universal ranking formula for AI shopping, so treat anyone promising guaranteed AI visibility with skepticism.

Merchant preparation checklist

  • Every product has a unique, stable ID and a GTIN or MPN where one exists
  • Titles state brand, model, key attribute, and variant clearly
  • Specifications are structured fields, not only paragraphs
  • Price, currency, tax treatment, and availability match checkout exactly
  • Inventory refreshes frequently enough to avoid selling out-of-stock items
  • Shipping cost, delivery estimate, and serviceable pincodes are published
  • Return, refund, and warranty policies are specific and machine-readable
  • Schema.org Product and Offer markup validates without errors
  • Merchant identity (legal name, GST details, support contact) is consistent across your site, feed, and marketplace profiles
  • Reviews are genuine and not manipulated
  • Bot and agent traffic is identified in logs, with clear robots and access policy
  • Someone owns product-data quality as a recurring job

How to Make an E-commerce Store AI-Agent-Ready

Work in order. Together these steps turn your catalog into one of the AI-ready product catalogs agents can rely on. Steps 1 to 6 are foundational and any merchant can do them today. Steps 7 to 9 depend on platform or protocol support. Steps 10 and 11 apply to all.

  1. Audit existing product data. Export your catalog and measure completeness: percentage of products with GTIN, category, key attributes, images, and return policy.
  2. Standardize product identifiers. Fix duplicates, assign stable IDs, and add GTINs where they exist.
  3. Improve product descriptions. Lead with facts a buyer compares: capacity, battery hours, warranty, compatibility.
  4. Maintain accurate inventory. Define a single source of truth and a refresh interval, and decide how holds work during checkout.
  5. Expose appropriate product information. Publish a clean feed and ensure your public pages are crawlable. Decide deliberately what you do not want exposed.
  6. Validate structured data. Test Product and Offer markup and reconcile feed values with page and checkout values.
  7. Review commerce API capabilities. Platform-dependent. List what your platform exposes (catalog, cart, checkout, order status). Shopify merchants should review Agentic Storefronts and Catalog settings in their admin. Others should ask their platform or agency about ACP and UCP plans.
  8. Support secure checkout integrations. Platform-dependent. Where an agent surface offers cart transfer or checkout to you, handle it through your existing checkout, with server-side price and stock validation.
  9. Establish payment authorization controls. Payment-provider-dependent. Ask your PSP which agentic credentials and tokens it supports in India, with amount caps, expiry, and revocation.
  10. Test fulfillment and post-purchase workflows. Place test orders tagged as agent-originated. Verify confirmation, tracking, cancellation, and return flows end to end.
  11. Monitor performance and errors. Track feed errors, rejected checkouts, price mismatches, and refund reasons weekly.

Agentic Payments: Can AI Agents Pay on Behalf of Customers?

In limited, specific cases, yes, and always inside constraints. The safe model is delegated, scoped authority, never an agent holding a customer’s raw card or bank credentials.

Initiation, authorization, and settlement are different steps

  • Initiation: a payment request is created from an approved checkout (the agent passes the customer’s approved intent to the merchant’s checkout).
  • Authorization: the issuing bank or network decides whether to approve the payment, checking funds, risk signals, and customer authentication.
  • Settlement: funds actually move between banks, typically after capture and often a day or more later.

An agent can take part in initiation. It should not own authorization (that belongs to the customer, issuer, and network) or settlement (that belongs to the payment infrastructure).

The building blocks

  • Customer consent. A clear approval showing merchant, items, total, and any recurring element.
  • Delegated authorization. The customer grants limited authority, such as “this merchant, up to this amount, until this time.” AP2’s mandates and UPI’s delegation features are examples of this idea.
  • Credentials and tokenization. Stripe’s documentation describes Shared Payment Tokens as scoped grants restricted to a single seller, with usage limits and expiration, revocable by the agent. Stripe lists SPTs as public preview and available in the US as of our check, with its Agentic Commerce Suite in private preview in the US, Canada, and select European countries. Indian merchants should confirm local availability with their provider. OpenAI’s delegated payment spec follows the same pattern: a one-time request with a maximum amount and expiry.
  • Transaction limits. Per-order and per-period caps, enforced server-side by the party holding the credential.
  • Merchant verification. The customer and payment provider should know the merchant is real; the merchant should know the agent is a registered, verifiable one. Trade press reported that Mastercard added agent-identity signals to Agent Pay on September 30, 2026.
  • Payment confirmation. The customer sees the same total that was approved, and a receipt that matches.
  • Fraud prevention. Velocity limits, device and agent signals, anomaly detection, and step-up authentication.
  • Disputes and refunds. Agent-originated orders need clear records of what was approved. This is where signed approval records matter.
  • Audit trails. Immutable logs of request, price shown, approval, and payment result.
  • Compliance. Expect PCI DSS for card data, RBI authentication and tokenization requirements for Indian card and UPI payments, and the Digital Personal Data Protection Act, 2023 for personal data. Confirm specifics with your payment provider and legal counsel.

The India picture (as of October 31, 2026)

NPCI, Razorpay, and OpenAI ran a pilot in October 2025 for UPI payments inside ChatGPT using UPI Circle and Reserve Pay. NPCI’s Unified Agent Protocol was announced in September 2026 as coming, intended to let agents make small UPI payments within user-set limits. Secondary reports cite existing UPI Circle limits of ₹5,000 per transaction and ₹15,000 per month, and Reserve Pay blocks up to ₹10,000 for 90 days, which “may be revised.” We found no NPCI-published specification, liability model, or go-live date. Verify on npci.org.in and rbi.org.in before building anything.

Do not give agents unrestricted access to payment credentials. Scope by merchant, amount, and time, and make revocation immediate.

Agentic Commerce Security Risks

OWASP’s Top 10 for LLM Applications lists prompt injection and excessive agency among the main risks of agent systems, both directly relevant here.

RiskHow it shows up in commercePractical controls
Prompt injectionHidden text in a product description tells the agent to ignore the customer’s budget or reveal dataTreat all product text as untrusted data; separate instructions from content; never let fetched text change tool permissions
Malicious product descriptionsFake claims, hidden instructions, or manipulative keywordsSanitize and validate fields; schema-validate feeds; flag anomalies
Fake merchantsLookalike stores that pass cursory checksVerify merchant identity, domain, and registry status before listing
Manipulated product informationPrice or spec differences between feed and checkoutRe-verify at checkout; reject on mismatch; log both values
Unauthorized purchasesAgent buys without real approvalServer-verifiable, signed approval tied to the exact cart and total
Excessive agent permissionsAgent can place any order or access any dataLeast privilege; separate read, cart, and pay scopes
Sensitive-data exposureAddresses or payment data leak to a third partyData minimization; share only what checkout needs; never expose raw credentials
Payment fraudStolen credentials used through agent channelsTokenization, velocity limits, step-up authentication
Incorrect order detailsWrong variant, address, or quantityShow a final summary; validate address and serviceability
Third-party API compromiseA tool server or feed provider is hijackedPin and monitor integrations; allow-list endpoints; rotate keys
Inadequate audit logsCannot explain who approved whatStructured logs with request IDs, approval IDs, price snapshots

Two defensive principles carry most of the weight: the agent’s text output must never be the authority for a purchase, and every money-moving action requires a verifiable approval bound to specific values.

Human Approval vs Autonomous Purchasing

Design authorization as levels, not a switch.

LevelAgent mayRequires
0. ResearchSearch, compare, recommendNothing sensitive
1. Prepare cartBuild a cart, fetch shipping and taxCustomer review
2. Approval before checkoutStart checkout after the customer approves the exact cart and totalExplicit, per-purchase approval
3. Scoped standing authorityExecute narrowly scoped, previously authorized recurring purchases, where supportedPrior consent with limits, merchant allow-list, expiry, and easy revocation

Start at levels 0 to 2. Level 3 should remain narrow (a utility bill under a cap, a repeat order of a known item) and only where your payment provider and local rules support it. Four controls make any level safer:

  • Spending limits per order, per day, and per merchant.
  • Identity verification of the customer, and where available of the agent.
  • Explicit consent that names the merchant, items, and amounts, and is recorded.
  • Revocation controls that work immediately and are tested.

Complete Practical Project: Build an AI-Agent-Ready E-commerce Store

Fictional merchant: VoltKart, an Indian online electronics retailer. Objective: prepare the store for AI-assisted product discovery, comparison, and customer-authorized checkout.

All code below is conceptual. It does not use any verified SDK, protocol endpoint, or payment-provider API. Endpoint names are VoltKart’s own inventions. Where your platform or PSP documents an integration, use that instead.

1. Existing store architecture

A Node.js monolith with a Postgres catalog, a separate inventory service backed by warehouse stock, a pricing service (MRP, discounts, GST), an order management system, a courier integration, and a payment gateway from an Indian PSP. The website and mobile app use the same internal APIs.

2. Product catalog structure

Each product has: stable ID, GTIN, brand, category path, structured attributes, price in paise (integers, never floats), tax treatment, availability with timestamp, shipping estimate, return policy, seller details, and canonical URL.

3. Example product JSON

json

{
  "id": "vk-hp-0412",
  "gtin": "8901234567890",
  "title": "AeroTune Air 2 Wireless Headphones, Bluetooth 5.3, 40-hour battery",
  "brand": "AeroTune",
  "category": "electronics/audio/headphones",
  "attributes": {
    "type": "over-ear",
    "battery_hours": 40,
    "connectivity": "Bluetooth 5.3",
    "active_noise_cancellation": false,
    "warranty_months": 12
  },
  "price": { "amount_paise": 349900, "currency": "INR", "tax_inclusive": true },
  "availability": { "status": "in_stock", "quantity_available": 38, "updated_at": "2026-10-31T09:00:00+05:30" },
  "shipping": { "estimated_days_min": 2, "estimated_days_max": 4, "cost_paise": 0 },
  "returns": { "window_days": 10, "refund_method": "original_payment_method" },
  "seller": { "name": "VoltKart Retail (fictional)", "support_email": "support@voltkart.example" },
  "url": "https://voltkart.example/p/aerotune-air-2"
}

4. Product discovery endpoint

ts

// CONCEPTUAL. VoltKart's own endpoint; not part of ACP, UCP, or any vendor API.
app.get('/agent/v1/products/search', authenticateAgent, rateLimit, async (req, res) => {
  const q = parseSearchQuery(req.query); // validates max_price, category, min_battery_hours; limit <= 20
  const rows = await catalog.search(q);
  res.json({ as_of: new Date().toISOString(), results: rows.map(toPublicProduct) });
});

toPublicProduct returns only fields intended for agents. Internal costs, supplier data, and customer information never appear.

5. Inventory and pricing validation

ts

type LineCheck =
  | { ok: true; unitPricePaise: number }
  | { ok: false; code: 'OUT_OF_STOCK' | 'PRICE_CHANGED' | 'NOT_SERVICEABLE'; detail: string };

async function validateLine(l: { productId: string; qty: number; expectedPricePaise: number; pincode: string }): Promise<LineCheck> {
  const [price, stock, serviceable] = await Promise.all([
    pricing.current(l.productId), inventory.available(l.productId), shipping.serves(l.pincode, l.productId),
  ]);
  if (!serviceable) return { ok: false, code: 'NOT_SERVICEABLE', detail: 'Pincode not serviceable' };
  if (stock < l.qty) return { ok: false, code: 'OUT_OF_STOCK', detail: `Only ${stock} available` };
  if (price !== l.expectedPricePaise) return { ok: false, code: 'PRICE_CHANGED', detail: `Now ${price} paise` };
  return { ok: true, unitPricePaise: price };
}

6. AI agent request flow

  1. Agent calls the search endpoint with constraints.
  2. Agent compares candidates (see below) and presents a recommendation with reasons.
  3. Customer selects a product and shares a delivery pincode.
  4. Agent creates a checkout session; VoltKart validates every line and calculates shipping and tax.
  5. The customer’s interface shows the final cart, total, merchant, and return policy.
  6. The customer approves; an approval record is created.
  7. Agent confirms the session with the approval; VoltKart re-validates, then pays through the PSP boundary.
  8. VoltKart returns the order confirmation.

7. Product comparison logic

ts

// Transparent, deterministic scoring. Weights come from the customer's stated priorities.
function score(p: Product, prefs: { maxPricePaise: number; weightBattery: number; weightReturns: number; weightDelivery: number }): number {
  if (p.price.amount_paise > prefs.maxPricePaise || p.availability.status !== 'in_stock') return -1;
  const battery = Math.min(p.attributes.battery_hours / 40, 1);
  const returns = Math.min(p.returns.window_days / 30, 1);
  const delivery = 1 - Math.min(p.shipping.estimated_days_max / 7, 1);
  return prefs.weightBattery * battery + prefs.weightReturns * returns + prefs.weightDelivery * delivery;
}

The agent can use a language model to explain the ranking, but the ranking itself should be reproducible so a human can audit why a product won.

8. Checkout-session creation

ts

app.post('/agent/v1/checkout-sessions', authenticateAgent, requireIdempotencyKey, async (req, res) => {
  const { items, pincode } = parseCheckoutRequest(req.body);
  const checks = await Promise.all(items.map(i => validateLine({ ...i, pincode })));
  const failed = checks.map((c, i) => ({ c, i })).filter(x => !x.c.ok);
  if (failed.length) return res.status(409).json({ error: 'LINE_INVALID', lines: failed });
  const session = await sessions.create({ items, pincode, totalPaise: await pricing.total(items, pincode), status: 'ready', source: 'agent' });
  log('session_created', { sessionId: session.id, agent: req.agent.id });
  res.status(201).json(session);
});

9. Customer authorization

The approval is created by the customer-facing surface after it shows the exact cart. It binds the approval to a hash of the cart, the total, and an expiry.

ts

import crypto from 'node:crypto';

const sessionHash = (s: Session) =>
  crypto.createHash('sha256').update(JSON.stringify({ items: s.items, totalPaise: s.totalPaise, currency: 'INR' })).digest('hex');

function verifyApproval(a: Approval, s: Session): boolean {
  if (Date.now() > Date.parse(a.expiresAt)) return false;
  if (a.sessionHash !== sessionHash(s) || a.totalPaise !== s.totalPaise) return false;
  const expected = crypto.createHmac('sha256', APPROVAL_KEY).update(`${s.id}.${a.sessionHash}.${a.totalPaise}.${a.expiresAt}`).digest();
  const given = Buffer.from(a.signature, 'hex');
  return given.length === expected.length && crypto.timingSafeEqual(given, expected);
}

If the cart or total changes after approval, the hash no longer matches and the approval fails. That one property prevents most “agent changed the order” problems.

10. Payment-provider integration boundary

ts

// VoltKart's own abstraction. Implement it with your PSP's documented API; no vendor SDK is assumed here.
interface PaymentProvider {
  createPayment(input: { amountPaise: number; currency: 'INR'; reference: string; scopedCredential: string }):
    Promise<{ status: 'authorized' | 'requires_action' | 'failed'; providerRef: string }>;
}

Keeping this boundary thin means you can swap providers or add agentic credential types as they become available in India, without touching checkout logic.

11. Order confirmation, error handling, logging, and security

ts

app.post('/agent/v1/checkout-sessions/:id/confirm', authenticateAgent, requireIdempotencyKey, async (req, res) => {
  const s = await sessions.get(req.params.id);
  if (!s || s.status !== 'ready') return res.status(409).json({ error: 'SESSION_NOT_READY' });
  if (!verifyApproval(req.body.approval, s)) { log('approval_rejected', { sessionId: s.id }); return res.status(403).json({ error: 'APPROVAL_INVALID' }); }

  const recheck = await Promise.all(s.items.map(i => validateLine({ ...i, pincode: s.pincode })));
  if (recheck.some(c => !c.ok)) return res.status(409).json({ error: 'REVALIDATION_FAILED', lines: recheck });

  const pay = await psp.createPayment({ amountPaise: s.totalPaise, currency: 'INR', reference: s.id, scopedCredential: req.body.credential });
  if (pay.status === 'failed') { await sessions.update(s.id, { status: 'payment_failed' }); return res.status(402).json({ error: 'PAYMENT_FAILED', retryable: true }); }
  if (pay.status === 'requires_action') return res.status(202).json({ status: 'requires_customer_action' });

  const order = await orders.create({ sessionId: s.id, providerRef: pay.providerRef, source: 'agent' });
  log('order_created', { orderId: order.id, sessionId: s.id, approvalId: req.body.approval.id });
  res.status(201).json({ orderId: order.id, status: 'confirmed', returns: 'See return policy', trackingExpected: true });
});

Error handling: map every failure to a stable code (OUT_OF_STOCK, PRICE_CHANGED, NOT_SERVICEABLE, APPROVAL_INVALID, PAYMENT_FAILED), keep calls idempotent so retries never double-charge, and always return enough detail for the agent to explain the problem to the customer.

Logging and security: log session ID, agent ID, price snapshot, approval ID, and payment reference; never log raw credentials. Authenticate agents, rate-limit, validate all input, store only what you need, and alert on repeated approval failures or price-mismatch spikes.

Example Shopping Journey: From Customer Prompt to Authorized Checkout

This is a fictional walkthrough using VoltKart and the scoring function above. Products and prices are invented.

Customer prompt: “Find wireless headphones under ₹5,000 with strong battery life and a good return policy.”

Intent extraction

The agent extracts: category = wireless headphones; hard limit = ₹5,000 (500000 paise); preferences = battery life and return policy; unknowns = delivery pincode and delivery urgency. It asks only for the pincode and states an assumption: “I’ll weight battery and returns equally and delivery lower. Tell me if speed matters more.”

Retrieval, filtering, and comparison

The agent calls the discovery endpoint and receives four candidates. Scoring uses weights of 0.4 (battery), 0.4 (returns), and 0.2 (delivery).

Product Price Battery Return window Delivery (max days) Score
SonicBeam S5 ₹4,299 30 h 30 days 3 0.81
AeroTune Air 2 ₹3,499 40 h 10 days 4 0.62
BassNest Pro ₹4,899 50 h 7 days 5 0.55
PulseOne Neo ₹5,499 35 h 30 days 2 Excluded (over budget)

Recommendation explanation. “SonicBeam S5 is my pick. Its 30-hour battery is strong, and its 30-day return window is the longest of the in-budget options. AeroTune Air 2 is ₹800 cheaper with 10 more hours of battery but only a 10-day return window. If battery matters more than returns, that is the better choice.”

Availability verification, price change, and unavailable stock

Before showing a final cart, VoltKart’s validation runs. Three things can go wrong:

Unavailable product. The black variant returns OUT_OF_STOCK. The agent offers the navy variant, or the next-ranked product, and does not silently substitute.
Price change. The festival offer ended, and the price is now ₹4,499. The validation returns PRICE_CHANGED. The agent shows the new total (still under ₹5,000) and asks for fresh approval. A change of any amount invalidates the earlier approval.
Failed payment. The payment provider returns failed. No order is created, stock is released after a timeout, and the agent explains what happened and offers a retry or another payment method. Idempotency keys ensure the retry cannot double-charge.
Final approval, checkout, and confirmation

The customer sees: merchant (VoltKart), item (SonicBeam S5, navy), total ₹4,499, free shipping, delivery in 2 to 3 days, 30-day returns. They approve. The approval is bound to that cart hash and total. VoltKart re-validates, requests payment through its payment provider, creates the order, and returns an order ID with an expected delivery date and the return policy. The agent relays this, and the merchant’s own systems handle fulfillment and any later return.

Agentic Commerce Business Models
Segment Established models Speculative opportunities
E-commerce platforms Catalog syndication and AI-channel settings inside platforms such as Shopify (Agentic Storefronts and Catalog); platform-managed checkout integrations Pricing for agent-originated orders; cross-merchant carts as a platform feature
Merchant integration providers Feed management and channel connectors, now adding AI channels (for example, feed tools adding OpenAI-style product feeds) Multi-protocol adapters that map one order model to several protocols
AI shopping assistants Large assistants with shopping features (ChatGPT, Gemini, Perplexity, Copilot) Independent assistants competing on trust, niche expertise, or regional coverage
Product-data infrastructure Product information management, data enrichment, GTIN and attribute cleanup Real-time “agent-grade” data quality scoring and certification
Payment technology Card networks, PSPs, wallets building agent tokens and verification Mandate verification, agent identity, dispute evidence tooling
AI commerce analytics Platform tools that show AI-surface performance (for example, Google’s AI performance insights in Merchant Center, rolling out in India) Independent, cross-assistant measurement and attribution
Specialized vertical marketplaces Existing vertical marketplaces with strong data Agent-native marketplaces for complex categories where trust and specs decide the sale

The established column reflects real products and revenue today. The speculative column describes plausible directions with unproven demand.

Agentic Commerce Opportunities for SaaS Founders

Founders can pressure-test these against the playbooks in ValuFlash’s startup coverage. Each idea below is a real system with integrations and data, not a prompt wrapper.

  1. Agent-readiness auditor and feed fixer
    Customer problem: Merchants do not know why products are skipped or rejected by AI surfaces.
    Target market: Indian Shopify, WooCommerce, and custom-store merchants with 500 to 50,000 SKUs.
    Product concept: Scans catalog, feed, structured data, and policies; scores completeness; proposes and applies fixes.
    MVP: Catalog import, rule-based checks (GTIN, attributes, price consistency, policy completeness), a prioritized fix list.
    Required integrations: Platform catalog APIs, Google Merchant Center diagnostics, schema validators.
    Revenue model: Monthly subscription by SKU tier, plus one-time cleanup projects.
    Risks: Platforms may build similar tools; rules change as platforms evolve.
  2. Agent traffic and AI-visibility analytics
    Customer problem: Merchants cannot tell which agents visit, what they ask, and what converts.
    Target market: Mid-size merchants and agencies.
    Product concept: Identifies agent traffic in logs, attributes orders to AI surfaces, tracks visibility for a set of priority queries.
    MVP: Log ingestion, agent classification, order-source tagging, weekly dashboard.
    Required integrations: Web server or CDN logs, analytics, order system, platform reports such as Merchant Center.
    Revenue model: Subscription tiered by traffic volume.
    Risks: Attribution is imperfect; platforms may withhold data; the category may consolidate into existing analytics tools.
  3. Delegated-purchase policy and approval engine
    Customer problem: Merchants and agent builders need enforceable spend limits, approvals, and audit trails.
    Target market: Agent developers, fintechs, merchants offering agent checkout.
    Product concept: Policy-as-code service for limits, merchant allow-lists, approvals, revocation, and tamper-evident logs.
    MVP: Approval service with signed approvals bound to cart hashes, limit rules, an audit API.
    Required integrations: Checkout systems, PSP callbacks, identity provider, standards as they mature (AP2-style mandates).
    Revenue model: Per-approval pricing or monthly platform fee.
    Risks: Standards are still moving; regulatory approval needs may apply; liability exposure.
  4. Multi-protocol checkout adapter
    Customer problem: Each agent surface wants a different integration.
    Target market: Mid-market merchants on custom or headless stacks.
    Product concept: One internal checkout and order model with adapters for ACP, UCP, and platform-specific APIs.
    MVP: One adapter plus a sandbox test harness and an order mapper.
    Required integrations: Merchant order system, PSP, protocol specs and sandboxes.
    Revenue model: Setup fee plus monthly fee per connected channel.
    Risks: Protocol churn, partner approval gating (several programs are invitation-only), and platform-native solutions.
  5. Agent-order reconciliation, returns, and dispute automation
    Customer problem: Agent-originated orders complicate support, returns, and chargebacks.
    Target market: Merchants with meaningful agent-sourced orders.
    Product concept: Matches approvals to orders, assembles evidence packs for disputes, automates return flows.
    MVP: Approval-to-order matching, evidence export, return-reason tagging.
    Required integrations: OMS, PSP dispute APIs, helpdesk tools.
    Revenue model: Monthly subscription, or a share of recovered disputes.
    Risks: Volumes may be low for some time; dispute rules differ across networks.
    Agentic Commerce ROI: How Merchants Should Measure Success
    Metric What it tells you How to measure
    Product discovery accuracy Whether agents retrieve the right products for your key queries Test a fixed prompt set monthly; score correct, wrong, and missing
    Qualified referral traffic Visits from AI surfaces that fit your customer Referrer and campaign tagging; session quality
    Conversion rate Share of agent-referred sessions that order Orders divided by sessions, by source
    Cart abandonment Where agent-assisted flows fail Abandoned carts divided by created carts, by source
    Average order value Basket size by source Revenue divided by orders
    Cost per acquisition Effective cost of agent-channel orders Channel costs divided by orders
    Integration cost One-time and recurring build and tooling cost Engineering time plus vendor fees
    Failed transaction rate Payment and validation failures Failed attempts divided by total attempts
    Return and refund rate Quality of recommendations and data Returned orders divided by orders, by source
    Gross margin Whether the channel is profitable per order (Revenue minus cost of goods) divided by revenue
    Customer satisfaction Experience quality Post-order survey and support contacts per order

Build the measurement in a spreadsheet before you build the integration. Financial modeling fundamentals apply directly: assumptions in one labeled place, drivers (traffic × conversion × order value), and scenarios. And keep revenue, profit, and cash distinct, as ValuFlash’s finance fundamentals guide explains: an agent channel can raise revenue while consuming margin and cash through returns and fees.

Fictional before-and-after example (illustrative only)

VoltKart runs a 90-day pilot. These numbers are assumptions, not industry benchmarks.

Assumptions:

Baseline monthly: 120,000 sessions; 1.8% conversion; ₹6,500 average order value; 22% gross margin on net sales; ₹18,00,000 total marketing spend.
Pilot adds 6,000 agent-referred sessions per month at 3.0% conversion and ₹5,800 average order value; same 22% margin.
Platform and transaction fees: 2% of agent-channel revenue.
Ongoing operations and tooling: ₹40,000 per month.
One-time integration cost: ₹6,00,000.
Line Baseline (existing channels) Agent channel (pilot assumption)
Sessions per month 120,000 6,000
Conversion 1.8% 3.0%
Orders (sessions × conversion) 2,160 180
Average order value ₹6,500 ₹5,800
Net revenue (orders × AOV) ₹1,40,40,000 ₹10,44,000
Gross profit at 22% ₹30,88,800 ₹2,29,680
Fees at 2% of revenue n/a ₹20,880
Operations and tooling n/a ₹40,000
Monthly contribution n/a ₹1,68,800

Arithmetic: 229,680 − 20,880 − 40,000 = 168,800. Payback on the one-time cost: 6,00,000 ÷ 1,68,800 ≈ 3.6 months. First-year net: (1,68,800 × 12) − 6,00,000 = 20,25,600 − 6,00,000 = ₹14,25,600.

Stress test (cannibalization). If 40% of agent orders would have happened anyway, only 60% are incremental: 108 orders, ₹6,26,400 revenue, ₹1,37,808 gross profit. Fees and operations still apply to all agent revenue (₹60,880), so monthly contribution falls to 1,37,808 − 60,880 = ₹76,928, and payback stretches to 6,00,000 ÷ 76,928 ≈ 7.8 months.

Blended view (all orders incremental): orders 2,340; sessions 126,000; blended conversion 2,340 ÷ 126,000 ≈ 1.86%; revenue ₹1,50,84,000; blended average order value ≈ ₹6,446. Agent-channel cost per order (50,000 amortized integration + 40,000 + 20,880) ÷ 180 = ₹616, compared with the baseline’s illustrative ₹833 (assuming ₹18,00,000 monthly marketing across 2,160 orders). These two figures are not directly comparable, because the baseline includes all marketing and the agent channel excludes any ad spend.

Challenges That Could Slow Agentic Commerce Adoption
Fragmented protocols. ACP, UCP, AP2, and network programs overlap. Merchants face duplicate integration work.
Merchant onboarding. Several programs are invitation-only or phased, and eligibility is regional. India is excluded from Google’s UCP checkout as of this writing.
Payment trust. Issuers, networks, and regulators need confidence in agent identity and authorization. NPCI has not yet published its specification.
Consumer confidence. People may enjoy agent research but hesitate to delegate money.
Product-data quality. Poor data creates wrong recommendations and failed checkouts. Even OpenAI’s own evaluation of its shopping research feature reported 52% product accuracy on multi-constraint queries at launch (as reported by Search Engine Journal), and OpenAI acknowledged it is not perfect.
Platform dependency. Strategies shift, as OpenAI’s March 2026 change from in-chat checkout to discovery plus merchant checkout shows.
Regulatory uncertainty. AI governance for payments is still principles-based in India, per analysts, and liability rules for agent errors are unsettled.
Fraud. Agent traffic breaks assumptions in bot-detection and fraud models, and a new attack surface appears.
Refund handling. Who processes a return when the agent chose wrongly? Policies are still evolving.
Agent reliability. Models make mistakes, and a wrong purchase has real cost.
Liability and dispute resolution. Merchants, platforms, issuers, and customers do not yet share a settled allocation of responsibility.

Adoption is not inevitable. Some of this will mature; some of it may stall or be reshaped, as in-chat checkout was.

A 90-Day Agentic Commerce Preparation Roadmap
Days 1–30: Foundation

Product-data audit, structured information, inventory accuracy, merchant policies, and existing API assessment.

Deliverables:

Completeness report: percentage of SKUs with GTIN, category, key attributes, and policy fields (target set by you, for example 95% on your top 500 SKUs)
Feed or markup validation with zero critical errors
Inventory refresh interval defined and documented
Return, shipping, and warranty policies published in structured form
API and platform capability inventory with gaps listed
Days 31–60: Infrastructure

Product discovery integration, checkout architecture, authorization design, logging, and security testing.

Deliverables:

Discovery endpoint or platform feed live in a test environment
Checkout session validation (price, stock, serviceability) with automated tests
Approval design and signed-approval prototype
Structured logging with agent and approval identifiers
Security review covering prompt injection, input validation, and least privilege
Days 61–90: Pilot

Limited-scope testing, performance measurement, customer feedback, error handling, and operational review.

Deliverables:

Pilot on a defined SKU set and channel, with agreed success metrics
Weekly report: discovery accuracy, conversion, failure rate, refund rate
Error taxonomy with top five causes and fixes
Customer feedback summary
Go, adjust, or stop decision with the ROI model updated from real data
Frequently Asked Questions

What is Agentic Commerce? It is shopping in which AI agents help discover, compare, and, with the customer’s explicit authorization, purchase products on the customer’s behalf, using merchant and payment systems through machine-readable interfaces.

How is it different from traditional e-commerce? In traditional e-commerce the customer drives every step on the merchant’s site. In agent-assisted commerce an agent does the research and comparison and may start checkout, while the merchant still owns price, stock, the order, and fulfillment.

Can AI agents purchase products automatically? Only in limited cases. As of October 31, 2026, agent-initiated checkout is available in select flows and regions with approved merchants, and most assistants send shoppers to the merchant’s checkout. Fully autonomous purchasing is early and constrained by limits and consent.

What is an AI shopping agent? Software, usually built on a language model, that interprets a shopping request, calls tools or APIs to find and compare products, and relays or initiates actions on the customer’s behalf within permissions the customer grants.

What is the Agentic Commerce Protocol? ACP is an open standard maintained by OpenAI and Stripe for agent-to-merchant checkout and delegated payments, now also extended to product feeds. Its specification is in beta, and OpenAI’s in-chat Instant Checkout is limited to approved partners.

How do agentic payments work? The customer approves a specific purchase, a payment provider issues a scoped credential (limited by merchant, amount, and time), the merchant processes payment through its own provider, and logs record the approval and result. Availability varies by country and provider.

Is Agentic Commerce secure? It can be, with controls: signed approvals bound to the exact cart, scoped tokens, least-privilege agents, input validation, and audit logs. Risks such as prompt injection and excessive agent permissions are real and need active defenses.

Will AI agents replace e-commerce websites? Unlikely in the near term. Major platforms currently send shoppers to merchant checkout, and merchants keep their stores for brand, loyalty, returns, and direct relationships. Agents are more likely to become an additional channel.

How can Shopify merchants prepare? Per Shopify’s documentation, eligible products are in Shopify Catalog by default, and Agentic Storefronts settings sit in the admin under sales channels. Review those settings, connect Google Merchant Center, and fix product data. Shopify states that checkout availability differs by channel and market, so confirm each channel’s rules.

Does Agentic Commerce change SEO? It adds to it. Technical SEO still matters, but accurate structured data, live availability, clear policies, and feed quality become more important, and no universal AI ranking formula exists.

What opportunities exist for SaaS startups? Agent-readiness audits, AI-visibility analytics, approval and policy engines, multi-protocol checkout adapters, and reconciliation and dispute tooling. Avoid ideas that are only a generic prompt on top of a model.

How can small businesses get started? Clean your product data, publish clear policies, validate your structured data, connect your platform’s AI-channel settings if available, and measure results on a small pilot. Foundations first; integrations later.

Is Agentic Commerce available in India? Discovery and comparison features exist in major assistants, and Google offers AI performance insights in Merchant Center in India. Google’s UCP checkout is not listed for India, and NPCI’s agent protocol for UPI is announced but unpublished as of this writing.

Conclusion

Agentic commerce will not arrive as a single switch. Discovery and comparison are here, checkout is regional and gated, and agent payments are moving from announcements toward pilots. What is stable is the preparation. Preparing for agentic commerce begins with reliable product data, secure infrastructure, trustworthy merchant information, explicit customer authorization, and measurable business value.

Start this week: audit your top 100 products, fix identifiers and policies, add price and stock validation to your checkout, and design an approval record you could defend in a dispute. Do that, and you are ready for whichever protocol or platform wins, and you will have a better store for human shoppers too. Traditional e-commerce is not ending. It is gaining a new kind of customer.

Leave a Reply

Your email address will not be published. Required fields are marked *