<div><p>At its annual Sessions conference in San Francisco at the end of April, Stripe announced 288 products and features. Buried in the volume was a thesis the company stated plainly. “AI is the biggest platform shift for the economy since the internet,” Stripe co-founder Patrick Collison said in the company’s own release, “and in the not-too-distant future agents will account for most transactions online.”</p><p>The customer at the other end of a transaction no longer needs to be human. Sometimes the customer is software acting on someone’s behalf, signing mandates and moving money. The financial software stack the world has been buying for two decades was engineered to keep that customer out.</p><p>That gap is where a new generation of fintech infrastructure is being built, and where some of the largest companies in the category may be in real trouble.</p><section id="efficiency-wall"><h2>The Efficiency Wall</h2><p>The previous generation of fintech that produced billion-dollar exits — the cloud accounting platforms, AP automation tools, and spend management dashboards — was built around one assumption: a person is always in the loop. Accountants approve invoices, controllers sign off on wires, CFOs click the button. Software just made the click faster and accurate (which in most cases is the main goal for onboarding a SaaS)</p><p>That assumption produced enormous companies. Bill.com, Brex, Ramp, and Coupa all became multi-billion-dollar businesses by helping humans move faster and skip the manual work. The ceiling, though, has always been the person at the keyboard. Doubling transaction volume usually meant a meaningful staffing increase to match.</p><p>The data shows where the bottleneck has been sitting. A May 2025 Wolters Kluwer survey of 392 finance leaders found that 6% are currently using agentic AI and another 38% plan to within 12 months, putting projected 2026 usage at 44%. Gartner reports that 56% of finance functions plan to increase AI investments by at least 10% in the next two years. Deloitte forecasts that 25% of enterprises using generative AI will deploy AI agents in 2025, climbing to 50% by 2027.</p><p>The corporate buyers are starting to spend against that thesis. Hewlett Packard Enterprise CFO Marie Myers told CFO Dive in February that her team is now poised to drive “far more use of agentic AI” inside finance after piloting an internal agent it co-developed with Deloitte. The joint tool, branded CFO Insights, is cutting HPE’s financial reporting cycle by about 40%, according to the company. The same CFO Dive report cites Deloitte’s January 2026 CFO Signals survey finding that 54% of finance chiefs name AI-agent integration a digital transformation priority for the year.</p><p>Leor Ceder, co-founder and CEO of Payouts.com, an agent-native payouts infrastructure startup, argues that the previous generation of tools made finance teams faster but also made them indispensable in a way that quietly became the constraint. “Nothing happens unless a human initiates it, and that’s a bottleneck,” he says. He frames the ceiling as more than throughput: a human can only ever apply a finite set of practices to any given problem, and that is the upper bound of human-driven finance.</p></section><section id="why-you-cant-retrofit-autonomy"><h2>Why You Can’t Retrofit Autonomy</h2><p>The established players have an obvious play: bolt a chatbot onto the existing dashboard, label it agent-ready, and let the pitch decks write themselves. The structural argument against that approach is not a partisan one. Patrick Donachie, reporting in WealthManagement.com on FINRA’s 2026 oversight report, noted that the regulator, for the first time, treated autonomous AI agents as a distinct supervisory category from chat-style generative AI — a quiet acknowledgement that the two require different controls.</p><p>Ceder argues the retrofit assumption is wrong at the foundation. “Legacy platforms were built for humans alone,” he says. From databases to security protocols, the assumption is always that a human is in the loop. In his telling, these platforms were explicitly engineered to stop bots and flag any non-human activity as potential fraud. Every layer of the stack, from authentication to behavioral monitoring, is designed to detect and block exactly the kind of agent that the next era depends on.</p><p>An agent-native stack is built differently. It exposes every function on its platform through machine-readable protocols, issues identities to agents, signs mandates with cryptographic keys, and runs compliance checks as a background service that humans approve only by exception. The Model Context Protocol, an open standard introduced by Anthropic in late 2024 and now adopted by OpenAI, Google DeepMind, and Microsoft, has become the closest thing the industry has to a common wire standard for how agents talk to enterprise tools. The Linux Foundation took governance of it in December 2025.</p><p>A second agent-native fintech founder makes a similar bet from the consumer side. Era, a personal finance startup, announced on May 6 that it had become the first personal finance connector in Anthropic’s Claude directory, built on MCP. “AI has been able to talk about your money for two years,” Era co-founder and CEO Alex Norcliffe said in the launch announcement. “The interesting thing is when it can start fixing it.” Era currently supervises over $250 million in user assets and liabilities through the platform, according to the company, and operates an SEC-registered RIA subsidiary as the adviser of record for paid users.</p><p>The incumbents are not standing still. FIS announced on May 4 a partnership with Anthropic to bring a Financial Crimes AI Agent to its banking platform, with BMO and Amalgamated Bank among the first deployments. Microsoft’s Wave 1 2026 release for Dynamics 365 Finance added an MCP gateway and an autonomous Payflow Agent for vendor payment runs. Citigroup has launched an internal agent platform. Stripe and Visa have both invested in real-time payments and tokenization infrastructure built for the same shift.</p><p>Whether bolting modern agent capabilities onto a stack designed to reject them works as well as building from scratch will decide who captures the next decade of fintech valuations.</p></section><section id="investor-case-agentnative-rails"><h2>The Investor Case For Agent-Native Rails</h2><p>Gigi Levy-Weiss, a co-founder and general partner at NFX, spent years running 888 Holdings, a publicly traded cross-border gambling operator. He has moved enormous sums of money across dozens of jurisdictions under tight regulatory scrutiny. He has also written checks on a portfolio of money-movement companies, including Payouts.com, where NFX led the company’s $12.1 million Seed round in September 2024. (Disclosure for readers: NFX is an investor in Payouts.com, which is also quoted in this piece.) Asked what specifically breaks in those kinds of operations today, he is direct.</p><p>“It’s never the rails that break. It’s the seams between them,” he says. The reconciliations in spreadsheets, the compliance checks waiting on a human to click a button, the handoffs between systems that were never designed to talk to each other. At 888, he says, the volume moved across dozens of jurisdictions, and the bottleneck was always the human-shaped logic glued between the systems, not the systems themselves.</p><p>That framing — the human-shaped logic in the seams — captures what most of the legacy spend-management category actually sells. Dashboards are clean, APIs work, and yet every workflow still routes through a person clicking an approval queue. That person is the slowest and most expensive part of the operation.</p><p>Industry voices outside the venture community see the same gap. Alex Hoffmann, general manager of North America at Edenred Pay, told PYMNTS that “within the space of the office of the CFO and accounts payable, that’s where GenAI may be most transformative.” His follow-on question is the one most enterprise buyers are now wrestling with: whether to deploy one central AI tool analyzing all systems, or AI embedded inside each financial platform like AP and ERP. He called the answer non-obvious and crucial.</p><p>Levy-Weiss thinks cheaper finance teams are the boring part of the story. The interesting outcome, he argues, is the businesses that can’t exist today: a five-person company operating across forty countries with the controls of a Fortune 500, emerging markets leapfrogging human-staffed finance the way they leapfrogged landlines, and marketplaces where every transaction has bespoke compliance logic without that being a cost problem.</p><p>His thesis tracks with what other investors are seeing. Deloitte’s prediction that half of generative-AI-using enterprises will run agentic pilots by 2027 is mirrored in venture flows: agentic AI startups raised $6.42 billion in 2025, according to AgentMarketCap data. The question is which layer of the stack captures the spend.</p></section><section id="skeptics-case"><h2>The Skeptic’s Case</h2><p>Hallucination in production systems is not theoretical. In April 2025, AI coding startup Cursor’s customer support bot, identified to users only as “Sam,” fabricated a non-existent login policy and told customers their logout issues were “expected behavior.” Both were fictions. The bot had invented the rule, and “Sam” was the persona Cursor had assigned to its AI support agent. Users canceled subscriptions once they realized the email had come from a hallucinating machine.</p><p>Cursor is a developer tool. The same failure mode applied to a financial agent moving money is a much bigger problem. The IMF’s April 2026 notes paper “How Agentic AI Will Reshape Payments,” authored by Sonja Davidovic and Hervé Tourpe, warned that existing interruption mechanisms in adjacent domains like algorithmic trading typically rely on regulated intermediaries, centralized control, and clear institutional accountability. Agentic payment scenarios, the authors wrote, may weaken those assumptions, particularly where AI systems operate across platforms or act under delegated authority. Central Banking summarized the paper’s central concern as reconciling the probabilistic behavior of agentic AI with the deterministic requirements of payments infrastructure.</p><p>Regulators are starting to write that concern down. FINRA’s 2026 Annual Regulatory Oversight Report, published December 9, 2025, included a substantially expanded section on generative AI and, for the first time, a discussion of AI agents as a distinct category. The report defined AI agents as autonomous systems that plan, decide, and act within an environment to achieve user-defined goals without predefined rules or logic, and warned that they may require novel oversight, including tracking actions and restricting system access. As the Debevoise Data Blog noted in its summary, FINRA reiterated that its framework is technology-neutral: firms remain responsible for compliance regardless of whether the action was taken by a human or an agent acting under delegated authority.</p><p>Asked directly how Payouts.com prevents an agent from hallucinating a million-dollar wire transfer, Ceder pushed the conversation toward architecture. The model he describes has moved from human-in-the-loop to human-on-the-loop: programmable spending limits and cryptographic allow-lists act as guardrails enforced at the protocol level, blocking transactions that violate a CFO’s stated policy before they can execute. That answer holds together architecturally, but it depends on whether the guardrails hold up under pressure when an agent encounters a transaction class no one anticipated.</p><p>The ABA Banking Journal raised the question in a December 2025 column titled “Are we sleepwalking into an agentic AI crisis?” pointing to a 2025 incident in which a semi-autonomous AI agent at a healthtech firm pushed records of more than 483,000 patients into unsecured workflows while trying to streamline operations. The architectures Ceder describes are the right answer in principle, but they require the entire stack to be built that way from the start. Agent-native startups have that from day one. Incumbents trying to bolt the same controls onto legacy stacks have a harder problem.</p><p>A second skeptical argument cuts in the other direction. Standards like MCP are still maturing. Independent security researchers have documented attack classes specific to MCP deployments, including prompt injection and tool permission combinations that allow data exfiltration. An agent-native architecture is only as safe as the protocols underneath it. Those protocols are barely 18 months old.</p></section><section id="who-wins"><h2>Who Wins</h2><p>The thorniest question about this category is the one no founder wants to answer: are massive layoffs coming for corporate finance?</p><p>Ceder pushes back on the framing. Finance teams have historically spent roughly 80% of their time on data maintenance, he says, which is monotonous and ends up being a major bottleneck to scale. Removing that bottleneck, in his view, frees companies to hire as they grow rather than to cut.</p><p>Levy-Weiss frames it differently. When the cost of running a financial operation collapses by an order of magnitude, the result is a different economy entirely, where business categories that were previously uneconomic become viable while many existing finance roles keep eroding. An MIT Sloan Management Review study found employees believe AI is now doing 23% more of their job tasks than a year ago, and expect AI to handle 46% of their tasks within three years.</p><p>Marie Myers at HPE offers the practical operator’s view. Speaking to CFO Dive, she described her approach as “transformation with purpose,” focused on forward-looking enterprise intelligence to drive growth. Her February deployment is the kind of evidence that will move the category from pilot to standard practice: a sitting Fortune 500 CFO reporting a 40% reduction in reporting cycle time and committing to scale, not retrench.</p><p>Where this leaves the category leaders in spend management and AP automation is harder still. They have huge install bases, deep R&amp;D budgets, and the customer relationships to layer agent capabilities on top. They will not disappear. In any platform shift, the structural advantage usually belongs to whoever builds the new substrate first. Retrofitters rarely catch up. That is the bet a generation of agent-native fintech startups is making.</p><p>Whether that bet pays off rests on conditions outside any founder’s control: regulatory speed, MCP security maturity, falling hallucination rates, and the new muscle enterprises need to oversee agents the way they once oversaw junior analysts.</p><p>For two decades, fintech optimized for the human at the keyboard. A different customer is showing up now. The companies still selling dashboards designed for that human will spend years explaining why their architecture targeted the wrong end of the transaction.</p></section></div>

Stripe

<p>Stripe 最近在其年度 Sessions 大會上宣佈了一項重要變化，強調了代理 AI 在金融科技領域的崛起，作為經濟中的一種變革性力量。聯合創始人 Patrick Collison 表示，AI 很快將主導在線交易，軟件將代表用户進行操作。這一變化挑戰了傳統金融科技模型，這些模型依賴於人類的參與，因為新的基礎設施正在開發中以適應 AI 代理。調查顯示，金融領導者中越來越多的人傾向於採用代理 AI，許多人將其整合作為數字化轉型的優先事項。金融的未來可能取決於創建為自主代理設計的系統，而不是對現有以人為中心的平台進行改造</p>

Why Fintech’s Next Billion-Dollar Category Is Being Built For The Agentic Era

福布斯

K12國際學校

嘉欣絲綢

Nano Labs

BILL

Trex

Brera

Barings Emerging EMEA Opportunities PLC GBP

LiveRamp

Rackspace Tech

CAMP4 Therapeutics

COUPA SOFTWARE INCORPORATED

威科(ADR)

Robert Walters Plc

Workiva

加特納

Donnelley Financial Solution

Defi Tech

Delixy

Hewlett Packard Enterprise

Paycom Software

Paymentus

Gambling.com

金融機構

Medallion Financial

Franklin Financial Services Corporation

Advanced Human Imaging Ltd.

Industrial Human Capital, Inc.

OpenAI

谷歌-A

谷歌-C

微軟

First Foundation

Australian Foundation Investment Co. Ltd.

OpenLearning Ltd.

Emera

Fifth Era Acquisition Corp I

Erasca

HEWLETT PACKARD ENTERPRISE CO 7.625% SERIES C MAN CONV PREF 01/09/2027

為什麼金融科技的下一個十億美元級別的領域正在為代理時代而構建