Commerce Without Clicks

When you open any product page on the internet, you'll see an image carousel designed for human eyes, a star rating designed to trigger social proof, and an original price crossed out next to a discounted one, designed to create urgency through the perception of a deal.

Now imagine you've deployed an AI agent to be the buyer on your behalf. It can't see images. It doesn't feel urgency. It doesn't have a browser session or cookies. Every single element of e-commerce was designed for a human, and every single one breaks when the buyer is software.

What's more is that the revenue model breaks when the shopper is an agent. Today, cart abandonment rates hover at 70% due to unexpected fees, confusing product options, and overplayed psychological games on the part of merchants. But in the future, when consumers adopt agentic shopping experiences, there will be increased pressure on merchants to change their own tactics. Nearly 60% of 18–34 year old shoppers say they trust an agent shopping on their behalf and, last year, 1 in 6 Black Friday purchases, $70 billion in GMV, were AI-assisted.

As agents become the primary shopper, the entire commerce stack needs to be rearchitected to fit their needs. The existing infrastructure that powers nearly $1.5 trillion in annual e-commerce sales in the US was built on a fundamental assumption that may become obsolete: that the buyer is a person.

The four-phase structure above captures the central insight: commerce doesn't flip from human to agent in one clean step. It splits. Humans keep the front end, the "I want winter boots" moment of intent formation. Agents run the middle, querying merchant APIs and executing programmatic checkout. Fulfillment, disputes, and loyalty stay human on paper, but agents co-pilot every piece: optimizing returns, assembling chargeback evidence, maximizing reward redemption.

The LLM layer is the bridge. When a consumer tells ChatGPT, Claude, Perplexity, or Gemini that they need winter boots, they're forming intent in natural language. The LLM interprets that intent, translates it into structured queries, and hands it off to an agent. This is where human discovery ends and agent search begins. The handoff is the most important architectural seam in the new commerce stack, and it's the least well-defined.

Human discovery still matters, but its function changes. Marketing, SEO, personalization, social commerce, and ads all target the human before they reach the LLM. Instagram shopping, influencer content, and brand storytelling still drive intent formation. But once that intent crosses into the LLM layer, the merchant loses the ability to influence the human. The agent takes over, queries structured product APIs, evaluates price and offer feeds, negotiates with merchant agents, and either counter-offers or commits. Discovery has shifted from pull to push: search will die as the primary interface, and agents will monitor preferences continuously and surface options proactively.

And thus, whoever owns the agent will own the top-of-funnel. The new gatekeeper is an LLM with entirely different ranking logic and no established playbook for winning placement. Perplexity, Google, and OpenAI have all announced agentic shopping experiences in partnership with the likes of PayPal and Stripe to power payments, and with marketplaces for access to multiple merchants at once.

The post-purchase phase is where the human-agent collaboration is most nuanced. Agents can and will co-pilot logistics (optimizing carrier selection, tracking shipments, flagging delays), facilitate returns (auto-generating labels, optimizing types of refunds), and even assist with dispute resolution. And as we argue elsewhere in this piece, loyalty will shift toward influencing agents: when an agent controls payment routing, loyalty becomes a machine-readable variable in an optimization function, not an emotional relationship. What remains human is the final authority, the consumer who approves a purchase, decides whether to keep or return, and ultimately judges whether the experience earned repeat business.

The implication is structural: the commerce stack is no longer a linear funnel that one actor traverses. It's a relay race between humans and agents, with the LLM layer as the baton exchange. The companies that build the handoff infrastructure, the translation layer between human intent and agent execution, will own the most valuable real estate in the new stack.

Today, the battle for top-of-wallet is fought through loyalty. Chase, Capital One, and Amex all aim to own the wallet by having a consumer-facing shopping portal. These platforms aim to target a human who sees a badge, feels a nudge, and picks their card based on emotional affinity of scoring a deal and the dopamine of point accumulation.

However, when an agent controls the workflow, including payment routing, the entire incentive structure inverts. The agent doesn't care about your brand. It has no card affinity. It optimizes for the principal's utility function: lowest effective cost, best rewards yield, fastest fulfillment. Loyalty becomes a machine-readable parameter in an optimization function, not an emotional relationship. Your five-star badge is just a UI component.

Soon, the battle for top-of-wallet will play out across four fronts:

This extends to advertising: the battle for human attention to drive the top-of-funnel is more important than ever. Today, advertising is critical because it captures human attention across display ads, sponsored listings, social media content, and remarketing. Agents don't have attention to capture. The human-facing half of advertising, influencer marketing, brand storytelling, social commerce, still drives intent formation in the discovery phase and keeps working. But the downstream conversion infrastructure, the part that turns attention into transactions, moves to protocols. On the agent side, merchants compete not by buying impressions but by offering better terms: lower prices, higher commission to the agent's principal, better return policies, faster shipping. The "ad" for an agent becomes a machine-readable bid offer, and the companies that build the agent-facing equivalents, structured offer feeds, machine-readable loyalty APIs, and programmatic payment routing, will own the next era of commercial infrastructure.

For agents to participate in commerce at scale, they need to authenticate, carry reputation, hold payment credentials, and act within delegated authority. This is the connective tissue across the entire agentic commerce stack, and it's the hardest part to build.

Today, identity in commerce is fragmented and human-centric: usernames, passwords, session cookies, 2FA codes sent to phones. None of this works for software. An agent needs a portable identity object, something like a DID (decentralized identifier) combined with a delegation framework that specifies exactly what the agent can and cannot do on behalf of its principal.

A delegation directive includes: spending limits (per-transaction and cumulative), category restrictions, merchant whitelists, payment method preferences, negotiation boundaries, and time-based expiration. Get identity wrong and you get either paralyzed agents (too restrictive) or rogue agents (too permissive). The infrastructure for this is still being assembled.

This is also where the fraud problem rewires. Today, fraud is about stolen card numbers and identity theft. In an agentic world, fraud is about compromised agent credentials, unauthorized delegation, and impersonation of legitimate agents. The Know Your Agent (KYA) problem is real: new fraud and risk systems will need to identify agentic fraud and bad actors, evaluating human vs. agent intentionality rather than just behavioral patterns. Fraud detection needs to shift from behavioral analysis of human purchasing patterns to credential verification and authority chain validation. Inferring intent through agentic actions is the most complicated problem we face today. The answer is contextual, dynamic policy at checkout: fraud providers score each transaction in real time, rewarding good actors and restricting bad ones. The companies that build the trust and detection layer will sit at the center of agentic commerce.

Every participant in commerce is about to get a new job description. Merchants, payment providers, consumers, and LLMs each sit somewhere different in the value chain once agents take the middle, and each one is going to be paid, compensated, or squeezed differently than they were when the buyer was a browser.

For merchants, the transition is an additive burden, not a substitution. The old playbook doesn't go away, humans still form intent before the LLM layer picks it up, which means ads, brand storytelling, SEO, and conversion optimization still matter for staying in the consideration set. But now merchants have a second job which is to build an entirely new layer on top: structured product catalogs, machine-readable pricing APIs, agent identity integrations, programmatic checkout. The brand work feeds the intent funnel; the agent infrastructure captures the transaction. Skip the first and the human never prompts for you. Skip the second and the agent routes around you even when you're the best fit. It's almost as if you don't exist anymore. Merchants who treat agent-readiness as a Phase 2 project will discover the transaction was already done by the time they got there. Merchant moat will come from the compounding effect of data infrastructure plus brand equity, but the cost of funding both from the same budget, in a market where LLM intermediaries keep extracting a bigger cut of the value chain, is where merchants may struggle. And yet, merchants trying to evaluate the ROI of budget expansion will be left behind.

For payment providers and FIs, the threat is disintermediation. Today, top-of-wallet is won through emotional affinity, card design, loyalty programs, and checkout placement. When an agent selects the payment method, it evaluates every available payment method in the human's wallet in milliseconds and picks whichever one nets the most value per dollar after rewards, fees, and settlement speed. Card networks and issuers need to compete on agent-readable terms, including programmatic reward rates, real-time authorization APIs, dynamic incentive structures that agents can evaluate in milliseconds. The FIs that embed themselves into the agent's decision function survive. Those that depend on human habit face a slow death.

American Express recognizes its advantage here and might be the only one who can solve the cold start problem. Acting simultaneously as the issuer, the network, and the acquirer is structurally advantageous. Amex aims to empower the consumer and the merchant to embrace agentic shopping by launching its Agentic Commerce Experiences (ACE) developer kit along with Amex Agent Purchase Protection, the first major issuer pledge to cover losses when a registered AI agent makes an erroneous purchase. The closed-loop network lets them do what no intermediated payment network can: register agents, verify intent, pass tokenized credentials, see full cart context, adjudicate disputes, and stand behind the entire transaction with capital. They brought the partners that matter from day one: Delta, Expedia, and Hilton on the merchant side; Adyen, Fiserv, Stripe, and PayPal on the payments side. The moat isn't logistics, it's the vertical integration of the money itself. When the biggest unsolved mystery in agentic commerce is who is liable when the agent gets it wrong, the network that already underwrites its cardholders' purchases is structurally the default trusted counterparty. Amex has always offered a premium experience for a premium price; in an agentic world, that posture stops being a marketing promise and becomes the technical spec.

For humans, agentic shopping is leverage. An agent negotiating on your behalf across ten merchant APIs simultaneously is a fundamentally better buyer than a human browsing alone. Prices come down, comparison quality goes up, and the time cost of shopping collapses. But this requires trust that the agent is acting on behalf of the human's best interest. Humans delegate purchasing authority to software they don't fully understand, mediated by LLMs whose incentive structures are opaque. The principal-agent problem is real: does the agent optimize for the human's utility function or for the LLM platform's revenue model?

For the LLMs, OpenAI, Anthropic, Google, Perplexity, the opportunity is to become the new access point and distribution layer. They sit at the handoff point between human intent and agent execution. They interpret natural language queries into structured commerce actions. They are the new gatekeepers, and their business model will likely include some combination of referral fees, merchant placement premiums, and consumer subscription revenue. This is the Google AdWords moment for AI: whoever controls the translation layer between "I need winter boots" and "POST /api/v1/cart/commit" captures the most valuable position in the new stack.

Mobile wasn't just a new screen. It was a new channel that restructured everything behind it: payments (Apple Pay, tap-to-pay), the experience of logistics (real-time tracking, push notifications, same-day as the default expectation), marketing (app-install ads, location-aware offers), and identity (biometric auth). The delivery networks mostly already existed; mobile helped assemble them, made them more accessible, and increased the demand for an instant experience. Every layer of commerce infrastructure was rebuilt for mobile because the human's relationship to the transaction changed: from episodic desktop sessions to a persistent, embodied, always-on presence that could track a package from a pocket, pay with a thumbprint, and act on an offer wherever it reached them.

Agentic commerce requires active supply-side participation: structured product APIs, machine-readable pricing, agent identity integrations, programmatic checkout, agent-accessible offer feeds, real-time inventory. Without merchant and marketplace buy-in, agents have little to no inventory to query, so the supply side has to opt in, and marketplaces are best positioned to be first movers, providing access to aggregated, real-time inventory. Marketplaces have the largest inventory sets, and partnering with LLMs is a clear path to survival for those that haven't built much aside from inventory aggregation. There are three types of marketplaces that likely survive the new world order.

Not all marketplaces land the same way. Let's talk about Amazon, where we have the clearest example, and their moat cuts both ways. For consumers: the logistics network (same-day/next-day delivery that no one can replicate), Prime membership (subscription lock-in bundling video, music, grocery, and free shipping), and catalog breadth (multiple sellers resulting in competitive pricing) force the agent to consider Amazon as the best place to buy from even if the agent might find a better price on a merchant API elsewhere. The bundled value proposition wins when the consumer's utility function weights delivery speed and return ease. For sellers: FBA (Fulfillment By Amazon) hands them the infrastructure an agent prioritizes, fastest shipping, cheapest returns, verified inventory, which means FBA listings structurally win agent-mediated queries even when the seller hasn't lifted a finger on agent-readiness themselves. The algorithm running Amazon's Buy Box has employed the mechanism of agentic selection for the last 20 years; sellers who optimized for it (best pricing, fastest to fulfill, high customer satisfaction) are already optimized for the LLM-routed query.

Everyone else can't wait. For the long tail of merchants, D2C brands, and marketplaces, the moment to join the agentic revolution is now. Perplexity, Google, and OpenAI have all launched LLM-native shopping assistants, which are already evolving into their next form. In March 2026, OpenAI sunset its Instant Checkout feature, which had tried to own the full transaction inside ChatGPT. It didn't work: only about twelve Shopify merchants ever went live, and users consistently preferred to finish checkout on the retailer's own site. A specific architecture failed, the one where the LLM tried to be both the pipe and the toll booth. But the pivot is the point here. OpenAI didn't retreat from commerce; it decided to concentrate on discovery and redirected transactions to retailer-owned surfaces. Target, Sephora, Nordstrom, Lowe's, Best Buy, The Home Depot, and Wayfair are now integrated through OpenAI's Agentic Commerce Protocol, co-developed with Stripe. The merchants that aren't indexed by these systems within the next 12–18 months won't be indexed at all, because agents don't browse. Agents will query, and if you're not in the structured data layer, you don't exist.

Five segments of infrastructure still need to be built. Here's the map:

The playbook for agentic commerce isn't written yet. The protocols for how agents authenticate, negotiate, and transact are just being formed. The trust frameworks for delegated purchasing authority are still theoretical. The data infrastructure that lets a merchant serve both a human browsing on their phone and an agent querying via API is being built in real time. Every layer we've described in this piece, from structured product schemas to agent identity to programmatic payment routing to the LLM handoff layer, represents an infrastructure gap that needs to be filled.

This is not a single-product opportunity. It's a full stack rebuild. The merchant enablement layer, the FI infrastructure layer, the permissioning and fraud layer, the analytics layer, the developer tooling layer: each one needs its own defining company. Unlike mobile commerce, which evolved over a decade, agentic commerce is being pulled forward by LLM platforms that are already shipping features to hundreds of millions of users. The distribution is moving faster than the infrastructure: the moment to participate is now.

Today we are standing squarely in the cold-start problem. You have many people planting trees, but disregarding how to architect the full forest. Merchants won't build agent-ready infrastructure until agent-mediated demand materializes; agent-mediated demand won't materialize until merchants are agent-ready. The platforms that can break this deadlock are the ones that own the full stack, end to end: the model, the marketplace, the payment rails, the logistics. Alibaba's Taobao Instant Commerce is already running agentic commerce at scale because Alibaba owns every layer, from Qwen, their own LLM, to Taobao to Alipay. Amazon's Rufus is building the same model on top of FBA and Prime. American Express is doing the same from the payments side, using its closed-loop network to bundle issuer, acquirer, and merchant coverage under one launch. End-to-end players bootstrap the category because they don't need other players to participate. They generate their own demand, provide their own supply, and prove the behavior is real. LLMs have the clearest opportunity to build this complete lifecycle, but they must provide incentives for all the necessary players to participate, rather than charging fees from the get-go and scaring off merchants and consumers. After all, even a company like Apple had to burn dollars to get merchants and consumers on board with Apple Pay.

Commerce has always been won by accomplishing the same three things at once: Convenience, Experience, and Value. Without completing the trifecta, no platform or company can win. In the agentic world, Convenience moves from one-click to zero-click. Experience moves from browse-to-buy to describe-to-buy. Value gets optimized mathematically by an agent that doesn't forget the better coupon or the faster shipping option. The winners are the companies that can deliver all three, which is exactly what end-to-end players do by construction and what infrastructure players enable for everyone else.

If you're laying the foundations for the future of commerce, we want to talk.