TL;DR
- Agentic payments are structurally real but commercially pre-validation. The vast majority of agentic transactions to date are experimental. X402 volumes are around thirty thousand dollars a day, nowhere near validation of agents as economic actors.
- Rawitz sees agent-mediated volume reaching the billions in roughly three years, not the trillion-dollar projections in circulation, and even then the definitional question of what counts as "agentic" remains unresolved.
- Agents don't need stablecoins. Most merchants don't accept them. Crossmint's thesis is that agents need access to many payment methods: stablecoin wallets, cards, and local rails alike. The "cards can't do microtransactions" argument is overstated, as Mastercard is already rolling out a blockchain-based Agent Pay product.
- The unsolved layer is not the token but risk, liability, and identity. Existing protocols (X402, Mastercard's, others) are incomplete. Nobody yet knows what the KYA standards should be, or whether a merchant needs to identify the human behind an agent.
- Crypto's clearest oversolving is protocol and L1 proliferation. There is no such thing as an agent-native L1. The place crypto most oversolves is the sheer number of protocols that exist today, the consolidation thesis TBW has pushed all year.
Speakers
- Alex Rawitz, Crossmint, Head of AI
- Aleksandar Bukovski, The Big Whale, Host & moderator
The Demand Problem: Real but Nascent
The question Rawitz hears most from institutional clients is simply: where are agents that are going to pay for things? Belief in the eventual market is near-universal; actual demand is early. A little traction exists in agent-to-agent and agent-to-API flows, but no clear category has broken out. Consumer assistants buying groceries, business assistants handling procurement, fashion assistants buying clothes are all still open questions.
Pressed on hype versus reality, Rawitz was blunt: “almost all of it is overhyped, nothing has been proven out, and the sector remains pre-validation on agents as new economic actors. Beyond the technology gap, there is a behavioral one. Most people are not ready to delegate payments to an agent.”
The Trillion-Dollar Projection Is a Definitional Trap
Rawitz declined to endorse the one-trillion-in-three-years projections. His experience is that there is always a way to fit reality to a projection if you want to, by counting any large category of transactions as "agentic" because an AI touched part of the flow.
The harder issue is definitional. Today the frontier is an agent that confirms permission for a single purchase before executing; the world where an agent manages a budget without approving every purchase is a long way off. Does "money moved by agents" mean a human approves every purchase, agents run their own budgets, or agents assign budgets to sub-agents? Nobody really knows. From the low tens of thousands today, reaching the billions in three years would already be an insane growth rate.
Why Agents Don't Need Stablecoins
Rawitz's core argument: “most merchants don't accept stablecoins. Crossmint's working thesis is that agents will need access to many different payment methods, a stablecoin wallet, cards, or otherwise, and Crossmint's job is to provide all of them. The roadmap runs Visa and Mastercard first, Amex once ready, then a long tail of Venmo, Cash App, bank accounts, and local rails like Pix.”
The framing mirrors merchant adoption itself. Merchants are famously reluctant to change, and across nearly twenty years of the crypto industry most have not wanted to accept crypto, which is why Rawitz doesn't think stablecoins are necessary at the settlement layer. The merchant-friction points Aleksandar raised, wallets, accounting treatment, holding stablecoins on the balance sheet, reinforce the same conclusion.
On the microtransactions counter-argument, Rawitz pushed back directly. Payments companies will not simply roll over. In June Mastercard announced Agent Pay, a blockchain-based product, and the notion that card networks will die because they can't do microtransactions has been vastly overestimated.
The Real Unsolved Layer: Risk, Liability, and Identity
The card networks have launched agent-native protocols, but like X402 and MPP they are incomplete, and the questions around risk are not solved. Nobody yet knows what the KYA standards should be, or whether a merchant needs to identify the human behind an agent's purchase. Merchants will often want to, partly for risk, partly to own the customer relationship and market to the buyer later.
Identity granularity is itself unresolved. Do you identify the agent, the model it uses, or the harness it runs on? Verifiable credentials are one answer, and Mastercard has published a verifiable-intent spec, but Rawitz expects many protocols rather than one winner. Shopify's Universal Commerce Protocol, ACP, and others are all live, which is good for an aggregator like Crossmint that makes them accessible to developers.
On liability, Rawitz was candid that it is an open secret with no settled answer. An agent can go rogue and buy random things, buy the right item in the wrong quantity, fail to execute a critical purchase, or get the timing, shipping, or destination wrong, dozens of failure modes. Guardrails are being built into Visa and Mastercard products and by independent builders, but the more interesting category is startups trying to price the risk of an agent's malfunction. Guardrails alone are a relatively simple product; guardrails plus risk-pricing is the compelling problem, and there is no data to do it today.
Legal Personality and Where Liability Lands
A recurring client request is issuing a card directly to an agent, which is not legally possible, because cards must be issued to a human. A legal framework granting agents standing may emerge in a few years, likely starting in a small jurisdiction such as Bermuda or Malta, with others watching and adapting, as happened with crypto regulation. For now the workaround is issuing to the human and granting the agent access via protocol.
On accountability, nobody currently treats the agent as an independent actor; everything is acted upon by the human. Liability routes back to the human who supplied the prompt or intent that kicked off the action, which is what Mastercard's verifiable-intent spec is trying to standardize. And practical identification bites in disputes: if your agent wants to contest a charge, you have to identify yourself to resolve it.
Asked how long until liability formalizes, Rawitz resisted the premise that it ever fully settles. Disputes never really get settled, gray area is why lawyers exist, but the guardrails, risk-pricing, and liability patterns should mature to a more established form in roughly two to three years.
The Reliability Bar: Better Than Humans, Not Perfect
On trust and the "money demands 100% accuracy" objection, Rawitz reached for the autonomous-vehicle analogy. AVs don't have to be 100% safe, only better than human drivers. A business automating procurement doesn't need the agent to be perfect, it needs it better than the humans doing the job now, and convenience or cost advantages may justify accepting some reliability hit.
He was unsentimental about current model quality. The more you actually use any AI product, the more you notice the mistakes and hallucinations. This stuff is a fair way from being dialed in.
What Europe Should Do
Rawitz wouldn't give Europe a dramatically different spin than the US, but flagged PSD2 as an interesting development. It aims to make verifiable credentials and passkeys more uniformly adopted. Passkeys are a strong UX upgrade over 3DS, and critically an agent can't handle 3DS on its own, whereas a passkey served back to its human operator is far more feasible. Companies should plan roadmaps around it. Anything that improves user experience drives volume, and passkeys add security, improve UX, and streamline agent activity in one move.
On what Crossmint actually builds, Rawitz positioned the company as use-case agnostic. It is the plumbing: how an agent gets access to money and executes transactions. For trading agents, Crossmint supplies wallets and on-ramps while the agent calls the Robinhood or other markets API.
Where Crypto Is Oversolving
The session closed on TBW's consolidation thesis. Rawitz's clearest example of crypto oversolving: “the industry needs no new L1s, and there is no such thing as an agent-native L1.” He cited a figure of roughly $7.5bn raised by a wave of these protocols against negligible recent revenue. The place crypto most oversolves is the sheer number of protocols that exist today. His parting note of surprise was that people are still showing up with new L1s in 2026.
Aleksandar closed on the house view: the great consolidation is needed, and most unused chains and infrastructure layers need to disappear for the industry to focus on a handful of winners, the normal trajectory of any maturing industry.
%201.png)
%201.png)
%201.png)
