Every AI Recommendation Is a Transaction Waiting to Happen

Why closing the transaction is the hardest unsolved problem in agentic commerce

Agentic commerce is poised to have a huge impact on how consumers buy things. The demand is already there: 51% of consumers say they would be open to an AI agent completing purchases on their behalf.

What's far less clear is how those transactions will actually happen.

Right now, the industry is in a period of rapid experimentation. More than $202 billion was invested in AI last year, and companies across the ecosystem are racing to figure out how agents should discover products, interact with merchants, and complete purchases.

The results have been uneven.

In September 2025, for example, OpenAI announced Instant Checkout inside ChatGPT, enabling users to buy products directly within the chatbot. But the company has since scaled back those plans, choosing instead to route purchases through third-party apps.

The broader ecosystem is encountering similar friction. At a recent investor meeting, Shopify President Harley Finkelstein noted that despite Shopify powering millions of merchants, only a small handful are currently selling through AI-driven shopping tools. The challenge, he explained, isn't merchant demand — it's the maturity of the agent ecosystem itself.

At the same time, the number of potential AI commerce surfaces is expanding rapidly. Any surface where an AI makes a recommendation — whether that's a publisher, a loyalty app, a customer service agent, a travel concierge, an enterprise procurement tool, or a vertical SaaS platform — is a potential transaction surface. The moment of influence and the moment of purchase are converging across the entire web.

In other words, the opportunity is enormous, but the infrastructure is still unsettled.

Most of the industry's attention has focused on building new protocols and integration layers to support agent-driven checkout.

But one approach has received far less attention — even though it already works across the existing web: Browser-based agents.

While the ecosystem continues to debate new infrastructure, browser-based agents may offer the fastest path to making agentic commerce viable today.

It's a conviction CartAI has held since the beginning: the future of agentic commerce will not be winner-take-all, and the infrastructure built to support it shouldn't bet on one either. 

The architecture of agentic commerce that's still taking shape

The question now isn't whether agentic commerce will happen — it's how it will work. Several technical models are beginning to emerge, each offering a different way for AI agents to interact with commerce systems.

Much of the current experimentation is also being driven by large platform ecosystems — OpenAI, Google, and Shopify — each exploring how agents could transact within their own environments. 

Understanding Popular Agentic Commerce Models 

* Table covers primary models in active development. MCP server integrations — through which retailers can expose checkout capabilities directly to compatible AI agents — are not included. Merchant adoption remains early and commerce-specific tooling is still taking shape. 

Exploring the ACP and UCP Models

Protocol-based approaches like ACP and UCP represent meaningful progress toward enabling agent-driven commerce.

ACP and UCP offer several important advantages.

Speed

Protocol-based checkout is nearly instantaneous. An AI agent sends a structured API request to the merchant's backend and receives a checkout session containing line items, taxes, and shipping options in milliseconds. The consumer confirms the purchase and pays.

Structured data

Protocols provide clean, typed data. The merchant returns a JSON payload with exact pricing, inventory availability, and fulfillment options. There is little ambiguity about what the consumer is buying or what the total cost will be. The agent doesn't need to "read" a webpage and interpret what it sees.

Merchant control

Both ACP and UCP maintain the merchant as the Merchant of Record. The merchant controls pricing, accepts or declines orders, and handles fulfillment and returns within their existing systems. Other models of agentic commerce can preserve this principle as well — and they should. Maintaining merchant control of the transaction is essential to building a sustainable ecosystem.

These are real advantages. For merchants that implement these protocols, the checkout experience can be faster, cleaner, and more deterministic.

But there are also significant challenges slowing adoption today.

Technical requirements

Implementing ACP requires merchants to build REST endpoints for creating and updating checkout sessions, implement webhooks for order lifecycle events, integrate with a compatible payment service provider (Stripe's Shared Payment Token is the first), submit product feeds, and pass OpenAI's conformance checks before going live.

UCP introduces a similar level of complexity. Merchants must expose capabilities through APIs, MCP, or A2A integrations, connect their catalog through Google Merchant Center, and participate in an early access ecosystem that is still evolving.

Protocol fragmentation

The current landscape resembles an iPhone vs. Android moment. Implement ACP and your products can be purchased through ChatGPT. Implement UCP and they can be purchased through Google AI Mode and Gemini.

But what about every other AI surface?

What about vertical AI applications, publishers, creator platforms, cashback programs, or loyalty apps that want to enable AI-powered checkout? Do merchants implement a new protocol for each one?

Both ACP and UCP aim to be open and interoperable. In theory, merchants implement once and any compatible agent can transact. That's the vision. But the ecosystem is still very early. ACP remains in beta. UCP has only recently launched. And e-commerce history suggests that "implement once, sell everywhere" promises often prove more complex in practice.

Merchant adoption

Ultimately, protocol-based commerce depends on merchant implementation. Retailers must build and maintain the integrations before agents can transact through them.

Even if large retailers adopt these protocols, millions of smaller merchants may not. The long tail of commerce — independent retailers, Shopify stores, and niche brands — often lacks the engineering resources to implement new agent protocols quickly.

And for publishers, creators, or commerce apps that want to enable agentic transactions today, waiting for universal protocol adoption may not be an option.

Consumer trust

There is also a broader challenge that goes beyond infrastructure. Agentic commerce will only succeed if consumers trust AI to complete purchases on their behalf. Whether the transaction runs through a protocol or a browser agent, the consumer must feel confident they are receiving the right product at the right price — and that their payment information is secure.

Building that trust is an industry-wide challenge that extends far beyond the technical architecture.

Browser-based agents make agentic commerce possible today

Much of the conversation around agentic commerce focuses on building new infrastructure — new protocols, APIs, and integration layers designed specifically for AI agents.

Browser-based agents take a different approach.

Rather than requiring merchants to implement new systems, these agents operate within the infrastructure that already powers online commerce: the web itself.

Retail checkout flows were built for humans interacting through browsers. Product pages, shopping carts, address forms, payment fields, and confirmation pages all follow patterns that have been standardized across millions of websites over decades. A browser-based agent simply interacts with those existing systems.

Instead of calling a protocol endpoint, the agent navigates a webpage. Instead of receiving a structured API payload, it interprets the same information visible to human shoppers. This approach does not require merchants to modify their systems, implement new protocols, or expose new APIs. The agent operates entirely within the checkout flows that retailers already maintain.

No merchant integration required

One of the primary constraints of protocol-based approaches is merchant implementation. Retailers must build new endpoints, integrate with new payment frameworks, expose product data, and pass certification processes before agents can transact through those systems.

Browser-based agents remove that dependency. Because they operate directly through a merchant's existing website, no additional integration work is required. The checkout flow an AI agent uses is the same flow that human customers use.

This allows agentic commerce to function immediately across the existing web ecosystem, rather than waiting for new infrastructure to be implemented merchant by merchant.

Immediate coverage across the long tail of e-commerce

E-commerce is not dominated by a small number of retailers. It is composed of millions of merchants, many of whom operate on platforms like Shopify, WooCommerce, or custom storefronts.

Large retailers may eventually implement agentic commerce protocols. But the long tail of merchants often lacks the engineering resources to implement and maintain new integrations quickly.

If a human customer can successfully complete checkout on a website, a browser-based agent can interact with that same workflow. This provides immediate coverage across the full spectrum of online commerce, including merchants that may never implement agent-specific protocols.

Orders flow through existing merchant systems

Another advantage of browser-based transactions is that they remain native to the merchant's existing systems. When an agent completes a purchase through a retailer's website, the order flows through the same infrastructure used for any other customer transaction.

From the merchant's perspective, the transaction is indistinguishable from a standard customer order. This avoids the operational complexity that can arise when new transaction channels are introduced.

Bot blocking remains an unsolved challenge

Browser-based agents face a significant and largely unsolved obstacle: merchant bot detection. Retailers increasingly deploy sophisticated systems — Cloudflare, PerimeterX, DataDome, and others — specifically designed to identify and block non-human traffic. These systems analyze behavioral signals, browser fingerprints, and interaction patterns to distinguish automated agents from human shoppers. There is no consolidated solution that works reliably across all merchants, and success rates vary meaningfully depending on the retailer. Some merchants actively invest in making automation harder over time. For any browser-based commerce system, bot blocking is not an edge case to be engineered around — it is a persistent, evolving challenge that requires ongoing investment to manage.

Transaction speed is a real tradeoff

Protocol-based checkout is nearly instantaneous. A structured API call completes in milliseconds; the consumer confirms and pays. Browser-based agents operate on a fundamentally different timescale — completing a transaction typically takes anywhere from three to fifteen minutes, depending on the merchant's checkout flow, page load times, and how much verification the agent encounters along the way. For use cases where a consumer expects immediate confirmation, that gap matters. Background or deferred checkout — where the agent completes the purchase while the consumer moves on — can make the latency acceptable in many contexts. But it is a genuine limitation, and any honest evaluation of browser-based agents has to account for it. 

Agentic commerce won't have a single winner

If the current landscape feels fragmented, that's because it is. But fragmentation is typical of early ecosystems, where multiple technical models compete and evolve before a dominant architecture emerges. Agentic commerce is likely to follow that same path.

The future of AI-driven checkout will not be dominated by a single approach. Instead, it will be built on a hybrid infrastructure, where different transaction models coexist and are used depending on the merchant, the platform, and the context of the purchase.

Protocols will scale through platform ecosystems

Protocols like ACP and UCP will expand rapidly where platform support makes adoption easy. Shopify has already built native support for both protocols, which means merchants operating on Shopify — from small direct-to-consumer brands to large enterprise retailers — can enable protocol-based transactions without building custom integrations.

That distribution mechanism is powerful. When a commerce platform integrates protocol-based checkout directly into its infrastructure, adoption can scale quickly across thousands or millions of merchants.

For merchants operating within those ecosystems, the advantages are clear. Protocol-based transactions are fast, deterministic, and based on structured data exchanges. The agent communicates directly with the merchant's backend, eliminating many of the ambiguities that can occur when interacting with a visual interface.

These protocols also create natural distribution into emerging AI commerce surfaces — including ChatGPT, Google AI Mode, Gemini, and Copilot — which are rapidly becoming new entry points for product discovery and shopping.

Browser-based agents will enable universal commerce coverage

But platform-native protocols only solve part of the market.

Millions of merchants operate outside ecosystems that currently support ACP or UCP, including retailers running on WooCommerce, Magento, custom-built storefronts, and regional commerce platforms. Many of these merchants lack the engineering resources to implement new transaction protocols in the near term.

For this long tail of commerce, browser-based agents provide an immediate path to agentic checkout.

For publishers, creator platforms, and commerce applications that serve a diverse merchant base, this universal coverage is essential. Waiting for universal protocol adoption would significantly limit the range of merchants that could participate in agent-driven transactions.

Hybrid models will emerge

The limitations of both approaches are what make a hybrid architecture not just practical, but necessary.

Protocol-based checkout is fast and deterministic, but depends entirely on merchant implementation. Browser-based agents provide universal coverage, but face real constraints — bot blocking has no consolidated solution across merchants, and transaction times of three to fifteen minutes are a meaningful gap compared to instant protocol checkout. Neither model alone is sufficient.

The most practical architecture combines both. When a protocol endpoint is available, the system uses it — completing the transaction in milliseconds through a clean, structured integration. When one is not, the system falls back to browser-based checkout, accepting the tradeoffs in speed and reliability in exchange for coverage across merchants that may never implement ACP or UCP.

From the consumer's perspective, the experience remains as seamless as the underlying infrastructure allows. The agent selects the most efficient path available. Sometimes that means instant checkout. Sometimes it means a transaction that completes in the background while the consumer moves on. The goal is not uniformity — it is coverage.

This is what makes the hybrid model the only realistic path forward. Protocols will mature and merchant adoption will grow, but the long tail of commerce is large and slow-moving. A system that can only transact where protocols exist is a system that excludes most of the web. A system that relies exclusively on browser-based agents carries unnecessary latency and reliability risk where better infrastructure already exists. The right answer is to use both, and to be honest about what each one costs.

This is the architecture CartAI was built around. Rather than betting on protocols alone or browser automation alone, CartAI's infrastructure routes transactions through whichever path is most reliable for a given merchant — using protocol-based checkout where it's available, falling back to browser-based agents where it isn't. The goal is coverage across the broadest possible merchant base, not optimization for a single checkout model that doesn't yet exist at scale. 

Closing the commerce gap

Publishers, creators, and recommendation platforms already influence a large share of consumer purchasing decisions. What they have historically lacked is a reliable way to capture the transaction itself.

And the opportunity extends well beyond publishers and creators. Any platform where an AI makes a recommendation that precedes a purchase — loyalty and rewards apps, customer service agents, travel and concierge tools, enterprise procurement workflows, vertical SaaS platforms, financial services apps surfacing relevant offers — faces the same structural gap. The moment of influence and the moment of transaction are disconnected. Agentic commerce infrastructure is what closes that gap, regardless of the surface.

Agentic commerce has the potential to close that gap. But the infrastructure supporting it is still evolving, and universal adoption of new protocols will take time.

The current conversation around agentic checkout has become unnecessarily binary — protocols versus browser agents, structured versus unstructured, fast versus slow. In reality, different approaches solve different problems, and the right method often depends on the merchant, the surface, and where the ecosystem sits in its adoption curve.

Browser-based agents provide a practical path forward today. Because they operate through the existing web, they allow agent-driven checkout to work across the full range of merchants now — not just those that have implemented new integrations. But they come with real costs: bot blocking remains fragmented and unsolved at scale, and the speed gap versus protocol checkout is substantial.

Over time, protocols will mature and hybrid systems will emerge. But platforms that influence purchasing decisions cannot afford to wait for the entire ecosystem to standardize.

The goal is not to bet on a single approach. It is to meet the market where it is — and build infrastructure that works across the full range of merchants today, not just the ones that have already implemented the latest protocol. That's what CartAI is building: embedded checkout infrastructure for any platform where AI influences a purchase decision, across the open web, not just within a single ecosystem. 

There is no winner-take-all. There is only what works. 

Manil Uppal

Founder, CartAI