645 Thought Leadership

Agentic Commerce and Agentic Fintech: Where the Next Great Fintech Companies will be Built

McKinsey projects $3–5T in agentic commerce volume by 2030. The rails, protocols, and identity layers belong to incumbents — middleware and post-purchase are the two layers where a startup can actually win.

Agentic commerce has moved from concept to early production over the past eighteen months. OpenAI, Google, and Anthropic have all shipped competing transaction protocols, jockeying to become the standard upon which the future of commerce is built. Visa and Mastercard launched agent-ready card programs, and USDC settled $11.9T onchain in Q4 2025 alone, reflecting exponential growth and the fact that stablecoins are being used as core financial plumbing.

This report maps where the agentic shift creates venture-scale opportunity across commerce, payments, and financial infrastructure, based on 645's in-depth market research and discussions with experts. The takeaways below summarize implications for founders and investors; the full analysis follows.

Key Takeaways for Founders

Build Where the Gap is Structural. Middleware and post-purchase are the two layers where neither the card networks nor the LLM platforms can move fast enough to close the gap alone, which creates an exciting opportunity for startups. Rails, identity, and protocols belong to incumbents and standards bodies.
Plug into Existing Rails and Charge for What They Don't Do. Merchants adopt new payment rails slowly, and your take rate has to survive compression from the layers below you. The Stripe-Bridge deal at 0.3 to 0.5% net take rate is the benchmark.
Solve a Problem that Burns Today. For example, US retail returns run $890B a year, which clears that bar for a significant problem that can be the basis for a startup that reaches large size. A product that only works once agents transact at scale means betting on adoption timing as well as the product.
Innovate on the Business Model as Much as the Product. Robinhood grew through a business model innovation, reducing trading fees to zero. Chime did it with overdraft fees, and Nubank with the branch network itself. Idle-cash yield share is the agentic equivalent.

Key Takeaways for Investors

Respect the Sequence. Middleware and orchestration will produce the first large outcomes; AI-native financial institutions like Catena Labs carry the potential for the largest long-term exits. Size investment checks and entry prices accordingly.
Underwrite Founder-Market Fit in Regulated Domains. Founder-market fit is the single best predictor of durability across prior fintech waves. Agentic payments is dense with money transmission and chargeback liability obstacles, which puts a premium on founders who are agile and resilient enough to maneuver their companies around these challenges.
Track the Regulatory Gates. Section 1033, OCC guidance on third-party stablecoin yield, and FedNow adoption gate the self-driving money category. Investments made in this category bring opportunity but also regulatory risk.

Executive Summary

This report began as a look at agentic commerce and grew into an examination of how AI agents will reshape payments and the architecture of how money moves. The opportunity is a stack of interconnected layers, from infrastructure plays through full-stack AI-native financial institutions.

The near-term value sits in infrastructure. Consumer shopping apps like Daydream, Phia, and Wizard have generated attention and initial adoption but long-term traction and sustainability is still unclear. Our expert calls confirmed this view, and one expert put it plainly: the shopping platforms are not yet ready for prime time, due to underlying technology limitations preventing a seamless consumer purchasing experience.

The B2B side is further along. Commodity procurement, AP/AR automation, and back-office financial operations are where agents can deliver measurable ROI today.

The market is venture-scale. McKinsey projects global agentic commerce volume of $3T to $5T by 2030. By 2030 Bain estimates the US market alone at $300B to $500B, or 15 to 25% of US e-commerce, and in the same timeframe Morgan Stanley sizes US agentic GMV at $190B base and $385B bull. At a 1% blended take rate, the agentic payments infrastructure revenue pool reaches $30B+ globally by 2030.

Stablecoins are becoming the native payment rail for a machine-intermediated economy. USDC handled $11.9T in on-chain transaction volume in Q4 2025, up by 247%. This figure represents every transfer, payment, trade settlement, and movement between wallets recorded on-chain. Circle grew full-year revenue 64% to $2.7B. Stripe's Open Issuance prices at 1.5% versus 2.5 to 3% for cards. All of these metrics reflect stablecoins becoming widely-adopted.

Post-purchase is one of the clearest white spaces; self-driving money is the largest application-layer opportunity. US retail returns hit $890B in 2024 with no category leader in sight. New startups competing in this category include Siren. Self-driving money carries the biggest TAM but depends on open banking, instant settlement rails, and sustained agent reliability. Rivo is a 645 investment competing in the self-driving money category.

Adjacent agentic financial services are producing breakout outcomes on their own. Emerging categories within agentic financial services include AI banking (Catena Labs), agentic lending (Casca, Fuse, Salient), agentic accounting (Basis at $1.15B), and agentic investment banking (Rogo at $300M+ raised). We believe each of these categories will produce large exits. 645's portfolio company Uptiq is an emerging leader providing a suite of agents handling operational work for banks and credit unions.

How We Got Here

Each fintech wave has produced one or two middleware winners and several full-stack application winners. Plaid became the middleware layer of the cloud-and-mobile era, connecting 12,000 financial institutions to 8,000+ fintech apps. Visa tried to buy it for $5.3B before the DOJ blocked the deal. The company's revenue exceeded $500m last year.

The biggest fintech winners in the mobile wave were the application companies, which included Robinhood, Chime, Nubank, and SoFi. Nubank alone now generates $16.3B in revenue across 135M+ customers, and has a market cap over $60 billion.

Multiple lessons carry forward. First, infrastructure wins earlier, applications win much bigger, and founder-market fit in a regulated domain is the single best predictor of durability.

Second, the biggest winners changed the prevailing business model in a specific way, which was a critical source of their competitive advantage. Robinhood removed trading fees and account minimums. Chime dropped overdraft fees and made its money on Durbin interchange instead. Nubank skipped branches entirely and built around a no-fee card.

What the Agentic Wave Enables

The mobile and cloud waves changed how consumers reached financial services. Agents change who initiates and executes the transaction. On the commerce side, consumers will research, compare, and buy through agents instead of browsers; 20 to 40% of online shopping already involves an LLM somewhere in the research process.

For SMBs and the long tail of retailers, checkout, order management, customer service, and returns all become agent-driven. Programmatic B2B procurement is already working: Procure AI's Kärcher case study reports a 6% average discount and 90% supplier engagement on tail spend that was previously negotiated manually at marginal ROI. Procure's platform routes suitable purchase requests to agents based on criteria like commodity group and pre-set value thresholds, then negotiates discounts before any final purchase decisions are confirmed. The 90% supplier engagement rate indicates that suppliers will negotiate with an agent when friction is low; Procure engages with suppliers through existing channels.

On the financial services side, every consumer gets a personal financial manager, addressing pain points like idle cash optimization, bill negotiation, insurance shopping, tax filing. These services today require either a human advisor with a $250K minimum or a patchwork of apps. Agents close that gap for the median household, providing services that even a human advisor couldn't make. A human advisor rebalances quarterly; an agent rebalances daily based on rate changes, cash flow, and life events, and the same logic applies to bill pay, yield sweeping, and refinancing.

The Agentic Commerce Stack

Agentic commerce is a stack of six layers, and not every layer is investable for a startup.

Protocols and standards. Google (UCP), OpenAI/Stripe (ACP), and Anthropic (MCP) are competing. The protocols will be open; value accrues to tooling built around the winners. The layer is not venture-scale on its own, due to the business model and the presence of incumbents.

Payment rails. Existing infrastructure was built for human-initiated transactions. Agentic payments need programmable authorization, dynamic spending limits, and instant settlement. Visa and Mastercard are adapting with tokenized cards, and stablecoins are emerging as the parallel rail. This is the largest dollar TAM, a $30B to $50B infrastructure revenue pool by 2030, but incumbents hold the core.

Identity and security (KYA). When an agent hits a merchant site, someone has to verify that it is legitimate, whom it represents, and what it is allowed to spend. Our expert view is that the core identity layer sits with Visa and Mastercard; startups provide point solutions. Skyfire is an emerging player in this category.

Middleware and orchestration. Agent wallets, transaction orchestration, routing, and merchant acceptance. This is the best startup white space in the stack because it sits between the networks and the LLMs, and neither side can move fast enough to close the gap alone. We size it at $3B to $10B by 2030, derived as 10 to 20% capture of the $30B to $50B infrastructure revenue pool, itself a 1% take rate on McKinsey's commerce volume.

Merchant and catalog readiness. Agents need structured product data: exact specs, real-time inventory, accurate pricing, and return policies, while today's catalog data is built for human browsing. The category inherits the large existing Product Information Management (PIM) software pool, but Shopify and Salesforce may bundle this natively, so the window is open but uncertain.

Post-purchase infrastructure. Returns, chargebacks, loyalty, and disputes for agent-initiated transactions are largely unbuilt; every expert we spoke with acknowledged the gap. US retail returns hit $890B in 2024, 17% of merchandise sales, and no category leader has emerged.

Our thesis concentrates where startup right-to-win and near-term readiness are both high: middleware and orchestration, and post-purchase.

Where Incumbents Get Upended

Card networks preserve their position by adapting. Agents still need a settlement rail, and tokenized cards are the simplest path. Interchange survives, but agent-initiated payments create work the networks never had to do: issuing scoped credentials, verifying agent identity, and enforcing spend policies. Each is a product a startup can own, and together they form a take rate at the orchestration layer above the card that opens to new entrants.

Payment processors will be partially displaced; Stripe buying Bridge for $1.1B in October 2024 is the clearest precedent of a card-era processor buying its way into stablecoin rails. ACH faces the most pressure. It moves roughly $80T a year in batch-processed, human-initiated transactions, while agent-to-agent B2B needs instant settlement and programmable authorization with no human in the loop; stablecoins offer all of that natively.

Self-Driving Money

Self-driving money is an AI agent that manages your financial life autonomously. The wedge is idle cash: $5.93T sits in US checkable deposits per the Fed's Q4 2025 release, more than 5x the pre-COVID level. At prevailing rates, major banks earn roughly 4 points on those deposits while paying standard savings customers as little as 0.01%; the CFPB sued Capital One in January 2025 alleging $2B+ in lost consumer interest, and the New York AG refiled a similar suit after the CFPB dropped it.

Bank of America keeps its customers through high switching costs more than through high quality product. BofA and JPMorgan together added 28M consumer customers between 2020 and 2025, going from 129M to 157M, roughly 4% a year. Nubank and Revolut went from 44M to 199M over the same window, off products that pay real yield and charge no fees, and now count more customers than the two US giants combined.

An agent that moves idle cash automatically while offering personalized financial services attacks key weaknesses of US banks, providing startups with a window to move customers away from them.

Roboadvisors are the closest analog for what scale looks like: Vanguard Digital Advisor at roughly $300B in AUM, Betterment at $65B, and Wealthfront at $88B in customer assets and a $1.43B market cap post-IPO. Our bottom-up build puts the US SDM revenue pool at $3B to $12B by 2030, enough to support two to four venture-scale outcomes.

Adoption depends on four things going right. First, open banking: Section 1033 is finalized but under reconsideration at the CFPB, a genuine risk factor. Second, instant settlement: FedNow now has 1,400+ participating institutions. Third, yield clarity: the GENIUS Act bans stablecoin issuers from paying yield but is silent on third parties, and the OCC comment period closes May 1, 2026. And fourth, sustained agent reliability, because a single high-profile bill-payment failure sets the category back years.

What we are looking for: a wedge into idle cash before expanding to debt, taxes, or insurance; regulatory-aware architecture that avoids money transmitter classification, since 50-state licensing kills the economics; a product that sits on top of existing bank accounts, because the friction of switching primary banks is fatal; a fee structure tied to the yield the agent generates for the user; and founders credible on the regulatory side.

Rivo (645 portfolio company) and Piere are the two early movers in this category. JPMorgan could build this internally, but the classic innovator's dilemma applies: building it cannibalizes a business that runs on deposit inertia.

Middleware and Post-Purchase: Our Criteria

In middleware we want two-sided network potential, meaning the agent wallet and the merchant acceptance layer together. Sponge, a 645 portfolio company, is built this way. Startups also need to plug into the card networks or stablecoin rails that already exist, since merchants adopt new rails slowly.

Founder-market fit in payments or compliance also matters, since the domain is dense with money transmission and chargeback liability challenges. So does strategic distribution; Nekuda counts both Visa Ventures and Amex Ventures as investors, a signal that the networks may intend to partner with the orchestration layer rather than compete with it. Finally, the take rate has to be one the rails cannot eat. Middleware charges for work the networks do not perform, credentialing agents and enforcing spend limits, so its fee stacks on top of interchange rather than competing with it; Visa cannot compress a fee for a service Visa does not provide. Bridge, the stablecoin infrastructure company Stripe acquired, is the benchmark: it earns 0.3 to 0.5% net take as an orchestration layer above a settlement rail it does not own, the same shape of business.

In post-purchase we prioritize companies solving a hair-on-fire problem merchants already have, so the business works even if agent adoption comes slowly; US retail returns at $890B a year clears that bar. We are also prioritizing proprietary data generation, incentive alignment with the merchant (Loop and Narvar monetize return volume, which is a problem), a clear path from wedge to platform, and integration with the Shopify, Loop, and Gorgias workflows merchants already run. Siren is an emerging player in this category.

Adjacent AI-Native Financial Services

The same pattern is producing breakout outcomes across financial services: vertical AI agents in regulated, document-heavy domains are the durable winners. Catena Labs, founded by Circle co-founder Sean Neville, raised a $30M Series A in May 2026 ($48M total) to build the first regulated financial institution designed natively for AI agents. In lending, Casca ($29M, SBA origination), Fuse ($25M, credit unions), and Salient ($65M at a $500M valuation, $25M+ ARR, cash-flow positive) are the standouts. 645 portfolio company Uptiq differentiates itself by providing a full suite of agentic solutions for banks and credit unions covering digital banking, lending, account opening and wealth management.

Basis reached a $1.15B valuation in agentic accounting after completing a fully autonomous 1065 return, and Rogo raised a $160M Series D led by Kleiner Perkins for agentic investment banking with 250+ institutional clients including Rothschild, Jefferies, and Lazard.