Agentic Economy #08 | From Checkout to Choice: How AI Agents Are Reorganizing Commerce

The real opportunity in AI commerce starts long before the checkout button.

For the past three decades, ecommerce infrastructure has optimized for a single moment: reducing friction at checkout. One-click buying, tokenized credentials, Face ID, and fingerprint authentication all target the very bottom of the funnel—after the user has already made up their mind.

AI agents move the heavy lifting upstream. Before a payment is ever processed, an agent must understand user intent, compare products, build a cart, and execute choices within permitted boundaries. This makes agentic commerce far more than just a payments story. While payments still matter, they are now merely the final step in a completely reorganized flow.

Intent Becomes a Filter

Traditional ecommerce is a relentless browsing loop. Users search, open endless tabs, cross-reference reviews, jump between platforms, and get nudged by dynamic pricing, targeted ads, and algorithmic recommendations.

AI agents compress this entire loop. Today, a shopper might not even specify a platform or a brand. Instead, they simply describe what they want: a budget ceiling, a strict delivery deadline, brands to avoid, and a few highly specific preferences.

Previously, these preferences stayed locked in the user’s head. With an AI agent, they become programmatic filters. The agent’s job is to translate natural-language requests into actionable rules it can evaluate and execute.

This fundamentally shifts where value sits:

• The Old World: Merchants fought for clicks, impressions, and conversion on the product page.

• The New World: Products are screened before the human ever sees them.

While the product page still matters, backend fundamentals—like clean product data, live pricing, inventory accuracy, delivery promises, and structured return terms—will dictate whether a product even makes the agent's shortlist.

The Answer Becomes the Shelf

Once user intent is converted into constraints, those constraints feed directly into AI search engines and large language models. This has given rise to an entirely new product surface: the generated answer.

Traditional ecommerce started with a search results page. AI commerce starts with the final answer.

Where shoppers once waded through a chaotic mix of links, sponsored ads, and marketplace pages, generative search engines (like AI Overviews, ChatGPT, and Copilot) collapse multiple sources into a single, cohesive recommendation complete with use cases and buying advice.

This is why Generative Engine Optimization (GEO) feels far higher-stakes than traditional SEO:

• SEO was about visibility (fighting for a higher link).

• GEO is about judgment (fighting to be included in the model's neutral comparison at all).

The answer still comes from somewhere. Reviews, forums, short videos, comments, and buying guides all feed the system. A brand doesn't need an obvious AI ad slot to influence the result; it needs to shape what the model sees before it summarizes. One post looks like ordinary word of mouth. Repeated across enough surfaces, the same claim starts to look like evidence.

The Platform Dilemma

This shift makes platform governance incredibly difficult. In classic search, Google could display links and let users judge the credibility of the source. With AI Overviews, Google is generating the advice itself. If the answer is corrupted by fake reviews or content farms, the platform isn't just ranking a weak page—it's giving bad advice.

Different platforms will draw these trust boundaries differently. Microsoft may use GEO as an on-ramp for Copilot and Edge, while open-model platforms will have their own incentives. AI commerce won't have a single rulebook; it will have fragmented answer surfaces, each with its own commercial gates.

The Storefront Splits: Data for Machines, Taste for People

To be recommended by an AI agent, a product must first be machine-readable. Early ecommerce storefronts were built exclusively for human attention—using high-res imagery, emotional copy, and prominent buttons to prolong time-on-page.

Agents look past the aesthetic noise. They care about the underlying product record: SKU details, live inventory, net pricing, delivery windows, and verifiable return policies.

Because of this, machine readability is the new baseline for commerce. Technologies like schema markup, llms.txt files, and real-time inventory APIs will dictate an agent's understanding. While a model can scrape a standard webpage, scraped data is often stale or noisy. A structured catalog gives the agent a clean, high-confidence view of the transaction.

However, this shift won't hit every category at the same speed:

The Pre-Checkout Asset

Take the fashion app The Mall as an example. Online fashion discovery is notoriously scattered across Instagram, TikTok, newsletters, and creator feeds. The Mall pulls these fragments into a personal virtual space where users can track drops, follow brands, and use AI to seamlessly jump to similar styles across different labels.

This reveals the broader enterprise opportunity in agentic commerce: The pre-checkout layer—browsing, tracking, and comparing—is becoming a valuable standalone product. Managing fragmented taste sits incredibly close to consumer intent. The system that records what users save, ignore, or share across brands becomes the ultimate system of record for preference, holding data that may ultimately be worth as much as the final transaction commission.

Ultimately, the storefront splits in two:

1. The Machine Layer: Granular data, structured catalogs, and real-time APIs built for agent procurement.

2. The Human Layer: Brand expression, storytelling, and experiential curation built for human taste.

Moving the Promise Chain Upstream

Once an agent filters the options and builds the cart, the transaction finally hits the payment layer.

Traditional card networks are essentially delayed promise chains. At authorization, a merchant asks if the issuer stands behind the transaction. If yes, the goods move; clearing and settlement follow later. This framework assumed a human was always pulling the trigger and assuming liability.

AI agents disrupt this assumption. When a machine handles search, substitution, and cart execution, risk profiles change:

• What if the agent is manipulated by a prompt injection?

• What if it misinterprets a constraint?

The dispute is no longer just about whether a credit card was valid; it’s about whether the machine stayed within the boundaries of user intent.Some retailers are already trying to draw that line in their legal terms—Target, for instance, recently updated its terms to treat third-party agent purchases as explicitly user-authorized. While this kind of policy shift helps assign liability after the fact, it does not solve the underlying engineering problem. Agentic commerce fundamentally needs controls before the transaction happens.

To solve this, payment networks are building controls before authorization happens:

• Programmable Credentials: Visa and Mastercard are developing tokenized agent identities and single-use virtual cards limited by budget, merchant, category, or time window. If the agent goes out of bounds, the transaction is automatically blocked at the gateway.

• Developer-Layer Integration: Visa’s investment in Replit signals a move to embed these commerce protocols directly into the developer environment. By bringing the Trusted Agent Protocol into the build-and-deploy layer, payment networks want machine-native infrastructure to be intent-aware from day one, recognizing that future transactions will start inside custom developer software, not retailer apps.

The On-Chain Reality

Web3 and on-chain systems require an even earlier guardrail. Because a private key signature makes asset movement irreversible, constraints must exist before the signature is written.

This is the thesis behind contract-based agent wallets like Cobo Agentic Wallet and the Pacts underlying it. Instead of giving an agent open-ended wallet access, users issue temporary task contracts defining exact routes and spending limits. If a request breaches these rules, the MPC (Multi-Party Computation) nodes refuse to sign it, translating raw data into human-readable intent before execution.

The paradigm shift is clear: We are moving from trusting the agent to constraining the agent.

The New Architecture of Responsibility

Technology frequently changes the medium of commerce, but it rarely eliminates the need for accountability.

Ecommerce brought storefronts online. Mobile wallets changed the credential container. APIs made authorization programmable. Stablecoins are changing settlement. Yet, the core jobs remain identical: authorization, clearing, settlement, and disputes. Someone must always approve it, someone must stand behind the cash, and someone must take the blame when things go wrong.

AI agents stretch this chain to its limits. By outsourcing search, comparison, and checkout to background machines, the boundaries of user approval and merchant liability blur.

The infrastructure of the agentic economy must bridge this gap. We need a modern promise chain that binds user intent, agent authority, payment commitment, and dispute liability before the transaction ever occurs—making the entire flow completely verifiable and traceable.

AI commerce looks like an automation story. Under the surface, it’s a responsibility story.