Agentic Commerce: UCP, AP2, ACP, MCP and the Product Data Layer AI Agents Need

Agentic commerce is a new era of shopping in which autonomous AI agents act on behalf of users, handling everything from product discovery to purchase. If traditional e-commerce was about search and scroll, this new era is about ask and done.

To understand this shift, we need to look at the Digital Concierge concept. In agentic commerce, an AI can basically handle the whole shopping process for you: finding stuff, checking out options, and finishing the purchase with hardly any effort from your side.

So, picture this: you just tell your AI assistant, "Book me a nonstop flight to London next week for less than $600, and no overnight flights, please" (make sure you say "please" because you never know 🤫). The agent then goes and finds the perfect flight that meets the criteria, buys the ticket using your stored payment info, and sends the confirmation.

We're moving past the old way of e-commerce, where AI mostly just suggested stuff, to a totally proactive system where AI agents basically take the reins, planning, deciding, and making purchases for the user.

However, for this Digital Concierge to work at scale, we face a major technical challenge: How do we get millions of different online stores to speak the same language as these AI agents? The industry is solving this through a new stack of protocols.

The three layers below handle the buying: what gets sold (UCP), who is trusted to pay (AP2), and how the order is completed (ACP). One thing sits underneath all three: how an agent finds your products and understands what it is looking at in the first place.

That is the job of the Model Context Protocol (MCP). It lets an agent discover what a system can do and query it directly, without a developer hardcoding every call. MCP does not process payments, and it does not replace checkout protocols. UCP is built to work alongside it. If you want the developer view, see MCP vs. API and what a commerce MCP server does.

Imagine a travel agent forced to learn a completely new language for every single airline; they’d be paralyzed. That is the exact friction AI agents face today when attempting to parse mismatched product data across scattered e-commerce silos.

To take control of this fragmented landscape, Google launched the Universal Commerce Protocol (UCP) in collaboration with industry partners. Far from a peaceful consensus, UCP represents an aggressive play to establish a standardized set of functional primitives that dictate how AI shopping agents, retailers, and payment providers interoperate. Google positions UCP as an open-source, universal bridge designed to integrate with existing retail infrastructure, claiming compatibility with overlapping frameworks such as Agent2Agent (A2A), Agent Payments Protocol (AP2), and Anthropic’s Model Context Protocol (MCP).

But as tech giants wage this protocol turf war, merchants cannot afford to wait for a definitive victor. While empires fight over abstract standards, forward-thinking engineering teams are already deploying practical solutions such as the Crystallize MCP Server to expose their catalogs directly to models like Claude and ChatGPT, securing an immediate technical edge and real-world utility.

The underlying motive behind Google's protocol push is transaction capture. By enabling instant checkout directly within Google’s AI Mode in Search or the Gemini app, UCP lets agents complete transactions without the user ever stepping foot on a retailer’s website. While a merchant might snag a high-intent sale and reduce abandoned carts, they also surrender the storefront journey to the AI gatekeeper.

Furthermore, UCP’s ambitions have rapidly expanded from a simple checkout integration into a comprehensive commerce capability layer. The protocol aggressively targets five critical operational battlefields:

• Checkout: Initiating and finalizing purchases completely detached from the traditional storefront journey.
• Cart: Managing multi-item sessions and real-time quantity adjustments instead of rigid, one-off purchases.
• Catalog: Retrieving real-time product variations, pricing tiers, and direct inventory signals.
• Identity Linking: Exposing customer-specific loyalty perks, member pricing, and shipping advantages natively to the agent.
• Order Management: Providing post-purchase order tracking and transaction history downstream.

This completely upends traditional commerce architecture. Agentic commerce is no longer a bolt-on checkout widget; it is a high-stakes battle for real-time data orchestration across pricing, inventory, and identity. If these backend systems are fragmented, stale, or buried behind storefront-specific legacy code, the autonomous agent will simply skip your business to avoid making an unverified decision. Your glossy storefront means nothing to a machine; your only true defensive moat is providing structured, programmatic access to the absolute commercial truth behind it.

Where UCP works today. UCP powers checkout buttons on eligible product listings in Google's AI Mode and in Gemini. Google describes the rollout as phased, starting with select merchants through an early access program, and currently lists eligibility as the United States, Canada, and Australia. To take part, you need the checkout eligibility attribute set in your Merchant Center product data, complete and current feeds, a Google Pay and Wallet Console account, and a payment provider integrated with Google Pay. You stay the seller of record.

If you sell in the Nordics, treat UCP checkout as something to prepare for rather than switch on. The preparation is the same as for every other protocol: complete, accurate product data.

Once your agent can "read" the store (thanks to UCP), the next big issue is trust. If you tell an AI to buy something while you're offline, how does the merchant know that the AI is actually allowed to spend your cash?

To bridge this technical trust gap, Google introduced the Agent Payments Protocol (AP2) in 2025 as an open, standardized, and payment-agnostic framework. Traditional e-commerce workflows assumed that a human user physically triggered checkout. The agentic economy, by contrast, demands stringent programmatic assurances of trust and authorization. Specifically, the AP2 framework is designed to manage three core risks:

• Authorization: Ensuring the AI agent possesses explicit, unprovoked user permission for a specific purchase.
• Authenticity: Allowing the merchant's backend to verify that the incoming agent request accurately reflects the consumer's genuine intent, thereby eliminating unauthorized or altered algorithmic orders.
• Accountability: Creating a transparent, legally binding audit trail to clearly assign responsibility in the event of transaction errors or fraud.

To solve the Trust Gap, AP2 uses cryptographically signed mandates as proof of the user’s instructions. For example, when you instruct an AI agent to buy something, your request is captured in an Intent Mandate (a signed digital contract detailing what you asked for). Once the agent selects the item(s) and you approve, a Cart Mandate is signed, locking in the exact items and price. Together, these create a tamper-proof record linking your intent to the final order, forming a verifiable audit trail from request to payment.

AP2 opens up cool new shopping experiences that just weren't possible before. Imagine telling the AI agent, I really want this winter jacket in green, and I'm okay with paying up to 20% extra for it. If it's out of stock, the agent will monitor inventory and price, then automatically seal the deal with a secure purchase the second that green one pops up within the price limit.

However, this "set-and-forget" model—whether applied to automated retail apparel or multi-provider flight and hotel orchestrations—places an immense burden on the merchant's underlying data accuracy. While AP2 guarantees wallet safety and prevents an agent from overspending, the payment protocol itself cannot prevent a transaction from collapsing or triggering a costly return-logistics loop if the merchant's internal PIM backend serves stale data, misrepresents variant specifications, or suffers from real-time inventory race conditions.

The final architectural hurdle is structural orchestration: bridging the operational gap between the chat interface, where the customer is actively engaging, and the merchant’s checkout system, where physical inventory is managed. This specific communication boundary is the domain of the Agentic Commerce Protocol (ACP). Originally introduced through a close collaboration between Stripe, OpenAI, and Meta, ACP lets users discover a product through ChatGPT and complete the purchase directly in the conversational thread. ACP's core philosophy is to define a secure, uniform way for AI models, consumers, and merchant systems to exchange data and finalize a transaction.

ACP covers more than the buy button. It defines agentic checkout sessions, cart and product feed handling, delegated payment tokens, delegated authentication, and order webhooks for confirmation, shipping, and refunds.

While protocols like AP2 are designed to support fully autonomous, offline delegation, ACP focuses on streamlining the immediate Human-in-the-Loop checkout experience. The workflow operates through a straightforward conversational interface:

• A customer asks an LLM for curated product suggestions, such as the "best running shoes under $100".
• The AI agent displays relevant matching items, and if the merchant's infrastructure supports instant checkout via ACP, an embedded "Buy" button appears natively in the chat.
• The consumer clicks the button to verify order details, shipping addresses, and payment tokens without ever leaving the chat.

Within this loop, the AI model serves purely as an intelligent broker, orchestrating the exchange of user intent and transaction payloads without processing funds or directly handling physical inventory fulfillment. Because the platform acts as a middleman rather than a distributor, merchants retain full control over their inventory rules, payment processors, and customer lifecycles, and remain the definitive merchant of record.

Platform advocates often point out that merchants don't need a massive backend overhaul to join this ecosystem; if a business already routes transactions through Stripe, enabling agentic checkout can be as simple as adding one line of code.

However, this creates a major architectural paradox that decision-makers must confront: while a payment connection can be reduced to a single line of code, that pipeline is useless if the underlying AI lacks clean, real-time data to recommend the right size, color, or variant configuration in the first place. A seamless checkout button cannot salvage an experience broken by catalog data hallucinations.

To bypass critical data bottlenecks without triggering an expensive database migration, you can audit legacy storefront databases for the "lazy data" bugs that disrupt AI parsing. Combining these insights with the Crystallize Flare AI allows you to programmatically transform messy descriptions and flat CSVs into a highly structured product universe operating on absolute commercial truth.

While protocols like UCP and ACP set the rules of the road for AI agents, they don't solve the underlying transaction logistics.

The biggest bottleneck for Agentic Commerce isn't the AI; it's the data.

For an autonomous agent to close the loop (search, compare, and buy), it needs to read product details, check real-time stock, and understand complex variations (for example, distinguishing a waterproof shell from an insulated jacket). Legacy monolithic commerce platforms, built primarily for human eyes and visual HTML pages, often trap this data in unstructured blobs that AI agents struggle to parse.

This is where headless commerce and Product Information Management (PIM) platforms like Crystallize become the essential engine of the Agentic economy. Because Crystallize is built for APIs and structured data rather than static pages, it is inherently Agent-Native.

Crystallize gives agents two ways in. The GraphQL API returns structured product data, and the Crystallize MCP Server lets models such as Claude and ChatGPT discover and query the catalog without custom integration work. The same data feeds your storefront, so the agent and the shopper see the same price, stock, and specifications.

This isn't just theory. Want to see this in action? Check out our case study on Composable Architecture and AI. It shows exactly how structured data gives AI the context it needs (totally beating the unstructured blob problem) and turns technical readiness into a real-world edge over the competition.

So, we have UCP, MCP, AP2, and ACP. In big tech idealized PowerPoint slides, UCP dictates transactions, MCP talks to systems, AP2 locks down the digital wallet, and ACP holds the consumer's hand in the chat window.

But this entire theoretical framework completely breaks down if the product universe behind your storefront is messy and fragmented. The real architectural friction isn't protocol compatibility; it's data integrity.

When an autonomous AI agent queries your storefront to buy a green jacket, it doesn't care about your beautiful Tailwind CSS gradients or your fancy homepage carousel. It wants raw, schema-compliant, high-fidelity commercial truth. If your real-time inventory is lagging or your product variants are buried inside a random text paragraph, the AI agent isn't going to try harder—it’s just going to bounce and give the revenue to your competitor.

And that is where Crystallize enters the chat.

Instead of duct-taping a brittle legacy monolith and praying an LLM can hallucinate its way through your catalog, Crystallize is fundamentally architected for APIs and structured data. It cuts through the frontend fluff to give autonomous agents exactly what they crave: blazing-fast, highly structured, machine-readable product universes.

Because in the agentic economy, humans aren't scrolling your website anymore; they're letting their code do the shopping. You can either spend the next two years waiting for tech empires to settle their protocol turf war, or you can build a backend moat that actually works today with Crystallize.

SCHEDULE A 1-on-1 DEMO to see how Crystallize can help you with agentic commerce. Or, why not START building for FREE.