Is There Any Opportunity for Startups in Agentic Payments and Commerce?

When OpenAI and Google brought checkout directly into ChatGPT and Gemini through integrations with Stripe, PayPal, and Shopify, the story seemed obvious: collapse the funnel where discovery, recommendation, and payment all happen in one place and render search, SEM, and DTC Shopify stores secondary.

But the more I’ve dug into tokenization, agent protocols, and merchant infrastructure, the more I believe the real outcome is the inverse: Hosted checkout doesn’t disintermediate the stack - it solidifies it. Google, Amazon, Shopify, Visa, and PayPal end up stronger, not weaker. What matters is not who controls the rails but how the edges of the system expand to enable new coordination between agents, merchants, and consumers.

There will be a large number of companies in the agentic payments and commerce graveyard because the value is not in reinventing checkout or wallets. It will accrue to incumbents that already control tokens, vaults, and merchant infrastructure. The real openings for startups are narrow but meaningful, focused on building the protocols, ledgers, post purchase systems, and agent to agent payment rails that this new world will ultimately require.

This dynamic will play out because trust remains the fundamental currency of payments. Agents may abstract away user interfaces, but they will not replace the institutions that already safeguard credentials, enforce chargebacks, and manage fraud at global scale. In effect, agents will function as “smart tokens” that route intent through the same trusted vaults and rails that power commerce today, further entrenching the incumbents that control them.

Commerce in an agentic world is only as viable as your ability to authenticate and fund intent. That is why companies with millions of customers and stored tokens have a clear structural advantage. Case in point: Perplexity teaming with PayPal or Claude integrating with Stripe, both activating millions of credentialed relationships at once.

Incumbent wallets like Shop Pay, Apple Pay, Google Pay, and Visa Token Services already control millions of tokenized identities, bound to devices and secured by FIDO. We don’t need a new “agentic” wallet because agents can simply tap into the ones consumers already trust. Agents can use token vault forwarding - that is, credentials that are stored once in a vault (Stripe, Braintree, VGS, Adyen) can be re-used across transactions, passed as tokens rather than raw credentials. This means discovery can happen in ChatGPT or Gemini, but settlement still flows through existing PSPs.

My POV: consumer-side innovation in agentic commerce will not come from reinventing wallets or checkout flows. Value will accrete to the incumbents that already control tokenized credentials, and to the vaults and PSPs that manage and route those credentials at scale. 

Merchants need more than consumers with wallets. They need infrastructure that allows agents to read product data, place orders, and reconcile payments in a structured way. The next layer of agentic commerce is therefore being built around merchant interfaces and ledgers, which form the foundations for how agents will transact at scale.

Model Context Protocol, or MCP, is starting to look like the standard way agents talk to merchant systems. Instead of scraping messy HTML, MCP lets agents pull clean data on catalogs, pricing, and inventory through structured APIs. OpenAI already supports it and Google’s Gemini is moving in the same direction. On top of that, Google has introduced something called the Agent Payments Protocol, or AP2, which layers payments on top of MCP so agents can actually move money across wallets and rails. Together, MCP and AP2 form the beginnings of an interoperability stack that lets agents, wallets, and merchants all transact on common standards.

For merchants, the current e-commerce stack is inadequate. Most web stores are optimized for human browsing, not machine transactions. Agents need structured product detail pages (i.e., ‘agentic PDPs’) that expose fields such as pricing, inventory status, and affiliate logic in agent-readable formats. They also need merchant systems to maintain agentic ledgers: canonical order records that can accept transactions initiated by agents, track fulfillment, and reconcile settlement. In addition, merchants require authentication layers designed for agents. This “Know Your Agent” concept parallels KYC in financial services and would allow networks such as Visa and PayPal to validate agent activity, apply transaction limits, and mitigate abuse.

Platforms like Shopify, BigCommerce, Magento, and Stripe are best positioned to deliver these capabilities. They already manage checkout APIs, merchant onboarding, and fulfillment systems. By extending this infrastructure into agent-ready endpoints and ledgers, they can act as the interface layer that allows agents to transact reliably at scale.

Most of the volume of agentic commerce will flow through incumbent wallets, PSPs, and rails regardless of where discovery and checkout happen. 

If both discovery and checkout happen inside an LLM, the model becomes the primary interface for commerce, even if the merchant of record and acquiring stack remain with existing platforms. This is what we’re seeing with OpenAI’s integration with Etsy and Shopify. Users can complete purchases entirely within ChatGPT, but settlement still flows through Stripe and the merchants’ existing rails. In the future, though, we could see LLMs act as true merchants of record (MoRs), similar to how Amazon Marketplace or Apple’s App Store operate today, intermediating the transaction, holding the buyer’s payment credentials, and remitting funds to sellers on their own schedules. That would give the LLM end-to-end control over the transaction flow, fees, and dispute management. If discovery happens in the LLM but checkout is redirected to a merchant site, token vault forwarding takes over, and value accrues to the PSPs. If discovery and checkout happen entirely on a merchant’s site or platform, that platform remains the MoR and the existing rails continue to carry the transaction.

This is why utilitarian shopping (I made up this term) is such a good fit for LLMs. Commoditized, high-intent purchases like dish soap, textbooks, or toilet paper do not require inspiration or brand storytelling. All that matters is price, speed, and convenience. ChatGPT can surface five options for the best Swiffer for my house, let me (or an agent acting on my behalf with pre-set rules) compare them, and drop a payment payload into checkout.

By contrast, experiential shopping (non-utilitarian) will remain largely outside the realm of LLMs. Categories like fashion, design, furniture, and luxury are deeply tied to self-expression and discovery. No one wants an LLM to hand them five generic jeans options when the thrill comes from browsing Revolve, scrolling Instagram, or exploring a favorite brand’s catalog. Shopping is really an activity in and of itself, and often is a deeply personal one that reflects who I am and how I want to show up in the world, the level of disposable income I have, my spending habits and prioritization, and my desires for things I want to have but can’t. History reinforces this split: bookstores, CD shops, and DVD rentals disappeared quickly once e-commerce made them more convenient, but identity-driven categories have proven much harder to displace.

Landing pages will not go away because people have brand loyalty and enjoy the browsing experience in certain categories. If I love Pottery Barn and want to buy all of my furniture there, then I want to browse their site for ideas. At the same time, we also have loyalty to convenience and user experience. I buy everything I can on Amazon because I know it will arrive quickly and that if anything goes wrong, I am covered.

Aggregator discovery platforms with strong personalization and event-based shopping are also interesting. In a perfect world, I would want all my shopping history, sizes, preferences, inspirations from creators I follow, and details about my personal life (like upcoming events, travel, or even where I live) in one place. That app could then surface every product across categories and personalize recommendations just for me. Still, I would not use it for everything. If I need a dress for a wedding, I might check there, or I might just look at a few creators on Instagram or go straight to a designer I like.

The point is that discovery will not collapse into a single channel. It will remain fragmented across brand sites, platforms like Amazon, and aggregators with deep personalization, with different contexts driving where people choose to shop. There may be opportunities around the edges to create better shopping experiences or new discovery surfaces, but the underlying payment rails will not change. The value will continue to accrue to the incumbents that already control wallets, vaults, and settlement infrastructure.

The reality is there’s almost no room for startups in consumer or merchant checkout. Incumbents already control wallets and credentials, and platforms like Shopify and Stripe will own agentic interfaces. Competing head-on in either layer is a path to the graveyard.

The real opportunity is in the edges that incumbents have not yet addressed, especially agent-to-agent commerce and payments. This is where there is no merchant of record, and agents exchange data, services, or actions directly with each other. Traditional rails like cards and ACH were never designed for this; their economics collapse on sub-cent transactions, their batching and settlement windows introduce too much latency, and their dispute-resolution processes assume human involvement. They simply cannot support the high-frequency, real-time microtransactions agents will generate. If those payments are batched and settled later, incumbents will still capture the economics. But if they happen in real time, it demands new protocols and ledgers built specifically for machine-to-machine payments. This is where startups can define the rails. 

There are other seams worth watching. One is post-purchase infrastructure. Fulfillment, returns, warranties, loyalty, and support remain fragmented and inconsistent. Agents could unify these flows by automatically tracking deliveries, initiating returns, and filing warranty claims, while merchants could consolidate logistics and support into a single agent-driven layer.

Another is trust, identity, and fraud. Incumbent rails were built for human transactions with clear dispute paths. Agentic payments will need new layers of identity, KYA (know-your-agent), fraud detection, and dispute resolution that work in real time and at machine speed. Agentic transactions will look like bot activity to current payment systems, which are designed to block or challenge them through 2FA. The entire flow will need to be re-architected to recognize and trust machine-originated payments. This is a natural opening for startups.

Beyond payments themselves, startups can build enterprise-grade agentic wallets for procurement and multi-party approvals, or middleware that “agentifies” existing merchant systems by exposing clean PDPs, catalog schemas, and agent-ready endpoints. They can develop orchestration layers for MCP and AP2 that manage disputes, failover, and interoperability, or create agentic search engines that rank merchants based on latency and reliability rather than SEO. And they can build the monitoring and security tools that harden MCP and AP2 against injection, leakage, and drift.

These are narrow but meaningful openings. The winners will not be the ones trying to replace PayPal or Shop Pay. They will be the ones that make agents interoperable with each other and with merchants, and that own the infrastructure around the transaction rather than the transaction itself.

What changes is not who controls the rails, but where the edges of the system open up. Payments themselves will continue to run on the same wallets, vaults, and PSPs that dominate today, and discovery will stay fragmented across LLMs, Amazon, brand sites, and aggregators. The real shifts are at the margins: agents will start to transact with each other directly, creating demand for real-time microtransaction protocols; merchants and consumers will expect unified post-purchase experiences; and interoperability layers will become essential so agents and merchants can speak the same language. The payments and commerce stack is not being rebuilt, it is being reinforced, and the meaningful shifts will happen only where agents meet and need new infrastructure to connect.