March 23, 2026

OpenAI's first real attempt at agentic checkout failed on every metric that matters. CNBC reports that OpenAI is pivoting from Instant Checkout to merchant-specific apps inside ChatGPT, with Etsy, Walmart, and Shopify building dedicated storefronts where ChatGPT handles conversational discovery while merchants retain control of pricing, inventory, and payment. The original Instant Checkout recorded roughly 1% conversion, over 77% cart abandonment, and onboarded only about 12 merchants, per CNBC. It also charged a 4% transaction fee that made the economics unattractive for sellers. Kiri Masters, in a detailed postmortem on LinkedIn, captures what the numbers miss. "I made a fully agentic purchase through ChatGPT. A lavender candle. The experience was fine. Not bad. Not broken. Just not meaningfully better than what I already do on Amazon or Target." She never went back.

The failure is instructive because it reveals where agentic commerce does and does not create value. Digital Commerce 360 reports that OpenAI is repositioning ACP as behind-the-scenes infrastructure rather than a direct consumer surface. Checkout is already a solved problem on the web and in mobile apps. The incremental friction reduction from moving it inside an LLM is negligible. What agents do better than any existing interface is compress the discovery and evaluation phase. Masters cites Raj De Datta on this point. The bigger prize is improving "seeker" experiences, not duplicating mature payment flows. Shopify's internal data, cited by Masters, shows a 15x increase in orders originating from AI agents between January 2025 and January 2026. That growth is coming from discovery, not from agents reinventing how credit cards work.

The gap between discovery traction and conversion remains stark. Masters cites Adobe Analytics data showing GenAI-driven traffic to US retail sites grew 805% year-over-year during the 2025 holiday period. Morgan Stanley research, also cited by Masters, puts AI-assisted product discovery at 39% of consumers. Yet ChatGPT referrals still account for less than 0.2% of e-commerce sessions and convert 86% worse than other affiliate traffic, per Kaiser and Schulze data cited in the same analysis. Masters frames this as the performance gap that defines 2026. Traffic is surging but converting poorly because merchant infrastructure was built for human browsers, not machine agents.

The merchant-app model OpenAI is now pursuing shifts power dynamics in a way the checkout experiment never did. Under Instant Checkout, OpenAI owned the transaction and charged 4% for the privilege. Under merchant apps, Etsy and Walmart own their catalogs, pricing, and checkout flows while ChatGPT provides the conversational surface. This architecture gives large retailers exactly what they wanted, a discovery channel they do not have to rebuild for, while keeping payment economics intact. The question is what it means for smaller merchants who will not get a dedicated ChatGPT storefront. Masters notes that Shopify is opening its agent infrastructure to non-Shopify merchants, suggesting the answer may be platform-mediated access rather than direct integration. Olin Moran, also in the Masters thread, warns against betting on any single agent ecosystem. "You need to own an agentic layer that sits between agents and your back-end stack." For mid-market retailers, that abstraction layer is now the strategic imperative.

The same infrastructure gap shows up in B2B, though with different constraints. Digital Commerce 360 reports that agentic commerce faces a reality check in B2B e-commerce where fewer than one-quarter of B2B suppliers currently use agentic AI technologies despite rising investment intentions. Orders involving dozens or hundreds of line items, negotiated pricing, credit terms, and ERP integration make full autonomy premature. The consensus from Digital Commerce 360 is that AI agents are more likely to become an embedded layer of automation within existing digital infrastructure than to replace traditional e-commerce platforms outright. Quote-to-order assistance, contract-aware search, and reorder recommendations are the use cases gaining traction. Wholesale platform replacement is not.

Trantor pushes further, arguing that the assisted phase is already giving way to agent-to-agent negotiation in narrow, policy-bounded domains. Buyer and seller systems interact directly within predefined rules, optimizing discounts and accelerating transactions with minimal human intervention. Forrester, as cited by Trantor, expects one-third of B2B payment workflows to leverage AI agents by end of 2026, with one in five sellers responding to AI-powered buyer agents with dynamically delivered counteroffers. That is an ambitious forecast. Whether it holds depends on whether organizations can define governance frameworks fast enough. Digital Commerce 360 notes that B2B organizations are still developing the governance structures to define how much decision authority agents should have over purchasing and finances. The human role, as Trantor frames it, shifts from execution to oversight. But the organizations that have not yet defined what their agents are authorized to approve cannot make that shift.

PYMNTS offers the sharpest framing of what agentic commerce can learn from B2B payments. In agentic commerce, payment becomes the final settlement step after extensive machine-driven evaluation, mirroring long-standing B2B purchasing workflows. As purchasing authority shifts to software, the differentiating value is no longer frictionless checkout but the ability to communicate and enforce policies, credit controls, and risk thresholds in formats agents can interpret. Winners will have structured catalogs, verifiable policies, and predictable fulfillment. OpenAI learned this the hard way. The lavender candle was not a failed transaction. It was a successful purchase that nobody repeated because the agent added no value the buyer could not get elsewhere. The merchants who thrive in the next phase will be the ones whose infrastructure gives agents something genuinely useful to work with before the buy button ever appears.