Agentic commerce stopped being a demo in late 2025. OpenAI launched Buy it in ChatGPT on the Agentic Commerce Protocol (ACP) it built with Stripe; Google announced the Universal Commerce Protocol (UCP) at NRF in January 2026; and McKinsey began estimating that agents could mediate three to five trillion dollars of consumer commerce by 2030. The checkout moment, the part where an agent turns a chosen product into a completed purchase, went from concept to shipping standard in about two quarters.
The retreat that followed was telling. OpenAI launched Instant Checkout with Etsy and said more than a million Shopify merchants would follow "very soon." Six months later it scaled back: a relatively small number of merchants had gone live, the feature "did not offer the level of flexibility that we aspire to provide," and the transaction moved back to the merchant's own checkout. This happened against the easiest possible case, pickable products with stable SKUs and an assembled cart. If in-chat checkout proved hard to scale even there, the cases where the buyer can't simply point and buy are even more challenging.
The trust, payment, and merchant-of-record problems that ACP and UCP target are real, they are shared across all of agentic commerce, and the industry is solving them from more than one direction. Those problems don't go away for configurable products; configurable purchases inherit every one of them, and then have to handle something the pickable-product case never raises. Configurable products are items that have to pass a rule system, satisfy compatibility constraints, or be derived from a job spec before they are orderable. These products are common across B2B purchasing and found in many B2C categories as well, yet the agentic commerce conversation has barely engaged with them. The hard part that's distinctive to them sits upstream of buying, in the configuration itself.
What the Commerce Protocols Were Designed to Solve
What the commerce protocols are good at is the pickable-product funnel, and they are genuinely good at it. They differ in how much of that funnel they cover — UCP spans catalog discovery, cart building, and checkout; ACP centers on the delegated transaction itself — but they share the assumption that matters here: a product is something the agent can find, identify, and select, because it already exists as a sellable unit. That is a legitimate hard problem, and these are reasonable answers to it. It is also exactly the assumption configurable products break: there is nothing to discover or select until the configuration converges into an orderable Unit. The Instant Checkout retreat is informative precisely because it doesn't refute any of that work. OpenAI stepped back from owning native checkout inside the conversation, but the spec has continued to evolve, and over the same window the discovery and distribution path expanded, with agent-powered checkout moving into mainstream search results. The pickable-product funnel is being actively built out from multiple directions at once. That maturation is what throws the configuration gap into relief: the more capable the pickable-product path becomes, the more visible it is that none of it reaches configurable products.
The same premise, that "trust to transact" is the binding constraint, also explains the public demo set: coffee, snacks, household essentials, and branded apparel. None of this is wrong; it's just not the same problem as configuration.
What Breaks When the Product Isn't Pickable
Configurable purchases break that premise in three ways, and the breaks show up regardless of whether the configuration layer is a commercial CPQ running guided selling or a bespoke rule engine deriving a BOM from a job spec.
First, what gets ordered isn't a product page. It's a configuration that has to satisfy a rule system before it's quotable. The SKU may or may not exist before convergence. Sometimes there's a base configurable product with options that resolve to a quote line, and sometimes the orderable items are derived from a job spec entirely. Either way, the agent can't operate on "this product" as a stable target the way it can with a pair of shoes.
Second, validation isn't optional. The rules that govern what's buildable are typically encoded in systems with decades of compatibility, manufacturability, regulatory, and pricing logic behind them. A configuration that violates those rules isn't a checkout warning. It's a hard stop, often with safety or contractual implications. An LLM that hallucinates a valid configuration is a liability, not a feature.
Third, configuration is structured decision-making, not item selection. Whether the flow is guided selling through a rule-mediated decision tree, BOM derivation from a job spec, or constraint propagation against a product model, the buyer is making bounded choices that interact in ways the catalog itself doesn't represent. A shopping agent built around discovery, comparison, and checkout has no role in that decision-making unless something else exposes the structure to it.
None of these are edge cases. They are the dominant shape of B2B purchasing and a substantial part of regulated consumer commerce.
The Configuration Layer Doesn't Disappear
For every configurable product an enterprise sells, something is already doing the configuring. The shape varies: commercial CPQ running guided selling, a constraint solver wrapped in an internal API, a takeoff system that derives a BOM from job parameters, a homegrown rule engine hardened over years of order fulfillment, or some hybrid of these. What's constant is the function. The configuration layer is the source of truth for what's buildable, at what price, and with what constraints.
That existing layer is the architectural constraint. The agent doesn't get to replace it, and shouldn't try. The rules are bound to manufacturing realities, contracted pricing, regulatory limits, and inventory positions the agent has no view into. An agent that bypasses the configurator to decide validity on its own breaks the operating model behind it.
B2B raises the difficulty rather than lowering it. McKinsey's recent automation-curve work notes that in consumer commerce delegation is personal, while in B2B it is institutional, bound by procurement policy, budget approval, and risk review. That institutional weight is not a payment concern. It shows up as constraints the agent has to respect while shepherding the configuration itself: the policy that rules an option out, the budget ceiling that bounds the decision, the approval that has to be reachable before a config can be committed. For configurable purchases, that weight doesn't sit in the typical purchase path. It sits inside the configuration conversation.
Continuous with Guided Selling, Not a Replacement For It
What an agent layer adds to that configuration conversation is continuous with guided selling, not a replacement for it. Guided selling has been the right answer to "walk a buyer through a constrained decision tree" for two decades. The agentic layer extends it at three points: the front, the back, and the middle of the conversation.
At the front, the agent handles under-specified natural-language intent. A buyer might describe a goal or a constraint in their own words, and the agent translates that into a starting position the configurator's input schema can accept. The translation is real adapter-design work. The agent has to know the input model well enough to map intent into it, which means somebody has to expose that schema in a form the agent can use.
At the back, the agent turns the configurator's native constraint output into explanations the buyer can act on. That native output is usually rule IDs, propagation traces, or solver conflicts. This is the part that bites hardest in bespoke configurators, where the rule layer was usually never designed for external explanation. The metadata required to reconstruct "why" often doesn't exist yet, which means the adapter work isn't just integration. It's partly a refactoring exercise inside the configurator itself.
Through the middle, the agent holds cross-step context that page-based forms don't carry. That context includes the procurement policy mentioned three turns ago, the budget ceiling impliedby an earlier answer, and the brand preference inferred from account history. None of these are catastrophic gaps in current guided selling. They're just things humans have had to keep in their heads.
Each of these asks the configurator to expose something it currently doesn't. The commerce protocols have made buying legible to agents. The configuration layer isn't legible to them yet — and making it so is where the real work begins.
Where This Series Goes Next
The commerce protocols shipping today solve the right problem for one shape of commerce and don't reach another. For pickable products, the trust and payment work is real and being built out from multiple directions. For configurable purchases, we still need that trust and payment layer, but the hard work is found in the conversation before it: intent capture, rule- mediated decision-making, constraint explanation, and a clean contract between the agent and the configuration layer that already exists.
Part 2 sketches the reference architecture: a layered approach that puts the agent on top of those existing systems through a normalized adapter contract handling schema-aware translation and constraint explanation. A primitive worth tracking on the way there is MCP Apps, which went live as the first official MCP extension on January 26, 2026, and offers a way to connect agent conversations to configurator UIs without giving up the validation authority that has to stay in the configurator.
B2B agentic commerce is coming. The question is whether it gets built on top of the configurators that already encode what's buildable, or around them as if they weren't there.
References
OpenAI. "Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol." September 29, 2025.
OpenAI. "Powering Product Discovery in ChatGPT." March 24, 2026.
CNBC. "OpenAI's first crack at online shopping stumbled. It's preparing for the next wave." March 20, 2026. (Source for the merchant-adoption figures, via Forrester analyst Emily Pfeiffer.)
Google. "Read Sundar Pichai's remarks at the 2026 National Retail Federation." January 11, 2026.
McKinsey & Company. "The automation curve in agentic commerce." January 28, 2026.
Model Context Protocol. "MCP Apps — Bringing UI Capabilities to MCP Clients." January 26, 2026.