"Money is a better moat than data," one of my embedded finance CEOs told me when I asked why his AI rebuild wasn't being eaten by every general agent on the market.
The line crystallizes why embedded financial services continue to thrive as AI disrupts many enterprise categories. The seesaw of software stocks reflects the market still trying to price what software is worth at steady state, and which subcategories AI will absorb and which it will compound. Embedded finance moves financial services closer to customers and into the software experiences they use every day. They are both software and financial services. The financial service component is why we believe embedded finance is on the compound side, and the reason is structural.
I spoke with founders from four of our embedded finance portfolio companies to test the argument against the operators: Jose Bethancourt, founder of Method Financial; Logan Henderson, founder of Gridline; Max Friedman, founder of Givebutter; and Joshua Silver, founder of Rainforest. Their answers converged on four moat layers: (1) the deterministic core, (2) rails ownership, (3) liability and risk management, and (4) trust. Each layer is a way of being in private territory that frontier models cannot reach. Each layer compounds rather than compresses under AI pressure.
One question going forward is how agent-mediated commerce impacts the ecosystem. That is the part of the AI shift most likely to threaten consumer-facing fintech. At the infrastructure layer, it is being built on two parallel paths. The card incumbents are extending their rails, while stablecoin issuers and on-chain protocols are building a parallel set of rails.
The Four Moat Layers
Determinism at the core
Jose Bethancourt, founder of Method, said the system must be deterministic at its core. Method knows this well. They are the infrastructure layer for consumer liability data and payments. Even state-of-the-art models cannot reliably parse a statement balance, and they routinely fabricate values when the data is ambiguous. The accuracy bar in finance is fifteen-nines, not three. Joshua Silver at Rainforest, a payfac-as-a-service platform built for vertical SaaS companies, described the same rule. AI is prohibited from core systems entirely, including payment processing and the billing engine, because QA overhead exceeds the efficiency gain.
Two operators in different parts of the fintech stack independently landed on the same architectural rule, which is structural. AI accelerates everything wrapped around the deterministic core. It cannot be the core.
Rails ownership
Logan Henderson at Gridline said it cleanly. AI is not eating fintech because durable fintech businesses are not just software. They are systems built around owning the assets, infrastructure, and data layers behind money movement, investments, and transactions. AI accelerates parts of that. It does not replace it.
Operating these rails requires more than integrations. While embedded fintech companies typically do not operate within the regulatory perimeter, preferring to partner with licensed providers and banks, they must maintain a posture as if they did. The partner bank requires the same level of compliance discipline and supervisory oversight from the fintech, or the relationship won't survive. Because AI lowers the cost of impersonation and fraud, gatekeepers are tightening access in response. Banks hit by AI-impersonation scams are restricting phone-based access and screening partners for AI exposure. Every defensive move raises the barrier for anyone outside the system. The walled gardens are tightening.
The acceleration is consistent across the portfolio. Rainforest, a payfac for vertical SaaS, was running billions in annualized payment volume. Method's commerce expansion is creating product surfaces that rely on AI to detect fraud and identity at checkout. Gridline's AltComply did not exist before AI made private-markets compliance work economical to automate. Joshua Silver framed it structurally: existing customer base, proprietary data, and AI innovation compound for operators already in the category.
Liability and risk management
Logan Henderson at Gridline said, "Liability ownership sounds like risk. It's actually leverage when executed correctly to build a durable moat. It converts cost-center pricing into insurance pricing and creates switching costs that no feature comparison can replicate."
Gridline serves the alternative asset investment management market, working with RIAs and GPs in private markets. That category sits in a layer of finance where pricing follows responsibility, not feature count. AI changes the cost side of that equation. A better model makes the work cheaper and faster to execute. The category's pricing logic remains unchanged. As Logan put it: "AI compresses workflow software because workflow software is priced on features. It expands the value of liability-anchored infrastructure because that infrastructure is priced on risk."
Trust
Max Friedman at Givebutter put the underlying mechanism sharply: money is a better moat than data. Sitting in the flow of funds is what makes you the central nervous system of a customer's relationship with money, not a tool on top of it. The flow is the mechanism. The trust that gets built inside it is the moat.
Trust is what allows a customer to store funds on Givebutter's platform and let the team operate on their behalf. That kind of trust is not assumed; it is earned. Givebutter built it over years of over-delivering for both the nonprofits raising and the donors giving. Competitors cannot replicate that by shipping a better feature.
Trust backed by funds is a close relationship. It compounds under AI pressure rather than commoditizing. The product can be rebuilt around an AI-native experience for nonprofits and donors precisely because the trust already exists. A 75+ NPS customer base becomes a launch surface that a competitor would have to rebuild from the ground up, starting with the trust itself.
AI commoditizes the surface layer. It does not commoditize trust earned through years of over-delivery. The operators inside that trust can rebuild the product. The ones outside have to earn the trust first.
Agent-as-Buyer: Who Builds the Rails for Agent-Mediated Commerce?
The most forward-looking question for embedded finance is what happens when agents replace humans as the primary interface for commerce. The conventional read is that this displaces the incumbents. The actual evidence is more nuanced. The incumbents are active and building agent infrastructure on top of their existing rails. A parallel set of rails is being built outside that perimeter at the same time. Both stories are real.
The incumbent rails path
Visa launched Visa Intelligent Commerce in early 2025 and the Trusted Agent Protocol in October 2025 with ten launch partners. Visa now has more than 100 partners across the agentic commerce network and 30 actively building in the VIC sandbox. Mastercard launched Agent Pay, built around Agentic Tokens, scoped credentials with programmable spend controls. The first agentic transaction on Mastercard's network occurred in Q4 2025, and the full US rollout was in November 2025.
Stripe is the most aggressive operator at this layer. Released the Agentic Commerce Protocol with OpenAI in September 2025. Shipped Shared Payment Tokens that let agents initiate payments without exposing card data. In March 2026, Stripe integrated SPTs with Visa Intelligent Commerce, Mastercard Agent Pay, Affirm, and Klarna. Stripe's positioning is explicit.
The architectural pattern is consistent in the incumbent path. Agents transact using tokenized credentials issued by the incumbents, settling on existing merchant acceptance and card rails.
The stablecoin path
A different set of rails is being built outside the card perimeter. x402, an HTTP-native stablecoin payment protocol contributed by Coinbase to the Linux Foundation in April 2026, has processed more than 119 million transactions on Base and 35 million on Solana, with roughly 600 million dollars in annualized volume. Circle launched the Circle Agent Stack in May 2026: agent wallets, an agent marketplace, and nanopayments that settle USDC transfers as small as one-millionth of a dollar. Stripe and Tempo published the Machine Payments Protocol in March 2026 as a broader machine-to-machine standard. Google introduced AP2 for delegated spending authorization.
The economic case for the parallel path is concrete. 76% of agent transactions fall below the 30-cent fixed fee for card payments. Card economics do not work for micropayments at the machine scale. Stablecoin economics do. Today, 99% of agent payments settle in USDC, with 90% running on Base. Real volume is still small. On-chain data shows x402 processes about 28,000 dollars in real daily volume, much of the rest testing. The architecture is in place. The volume is not yet.
The incumbents are hedging into the parallel path. Visa moved roughly $7 billion in annualized stablecoin volume through VisaNet by March 2026. Mastercard's June 2026 on-chain verification announcement extends its agent infrastructure into stablecoin settlement. The boundary between the two paths is porous.
Method's "Know Your Agent" framing fits underneath both. The same KYC and identity infrastructure Method runs for human consumers extends naturally to agent verification at the bank-connection layer, whether the settlement rail is a card or a stablecoin.
Which architecture wins which volume is the open question. Whichever path wins, the infrastructure runs through identity, fraud, liability, and risk management, and trust is the layer that the embedded finance operators already own.
The Thesis Is Holding
AI does not erase embedded finance moats. It clarifies which layer they were always sitting on. The deterministic core, rails ownership, liability, and risk management, and trust are the layers that hold today. They get stronger under AI pressure as the walled gardens tighten and the operators inside accelerate.
The seesaw of public software markets over the last six months has treated pricing as one trade. The sub-categories tell a different story. Embedded finance is the compound side. The operators inside it are accelerating, and the agent-mediated commerce that was supposed to displace them may be built with them, on rails they either own or are hedging into.