The Lyceum: Fintech Weekly — Mar 19, 2026

The Fed held rates again this week — and then said it projects only one cut for the rest of 2026. That "higher for longer" reality is now colliding with three simultaneous shifts: AI agents are getting their first real bank accounts and payment rails, stablecoins are quietly processing trillions while regulators race to write the rules, and the buy-now-pay-later companies that survived the rate-hike era are applying to become actual banks. Your money is moving into new hands. The interest rate on it isn't moving at all.

Most fintech lenders don't actually hold your money — they partner with a traditional bank that does, and pay for the privilege. Upstart just decided it's done sharing the economics.

The AI lending company is applying for a national bank charter. If approved, Upstart could take deposits directly from customers and fund its own loans, instead of relying on partner banks that take a cut of every transaction. The timing makes sense: Upstart's 2025 revenue hit $1.04 billion (up 64% year-over-year), it swung to $53.6 million in net income after years of losses, and it's diversified into auto, home equity, and small business lending.

The most interesting new product is "Cash Line" — a small revolving credit line of $200–$5,000 designed to compete with payday lenders. An AI lending company going after payday loans while simultaneously applying to become a federally chartered bank is either very ambitious or very telling about where the industry is heading. The OCC's response will be a signal for the entire category.

Passing America's first stablecoin law last July was the easy part. The GENIUS Act directed federal and state regulators to finalize rules on issuer licensing, capital requirements, custody, and anti-money laundering by July 18, 2026. That deadline is now four months away, and the biggest fight is over yield.

Here's the tension: banks make money by paying you low interest on deposits and lending that money out at higher rates. If a stablecoin — a digital dollar on a blockchain — can pay you a competitive return and settle payments instantly, the value proposition of a savings account starts to wobble. Banking groups are lobbying hard to close what they call a "loophole" allowing stablecoin yield. The crypto industry is fighting back.

States aren't waiting. Florida's SB 314 passed the legislature this week, creating a registration regime with full 1:1 reserve requirements — a model that could spawn a patchwork of state-level frameworks if the governor signs it. At the D.C. Blockchain Summit, key senators signaled renewed urgency on a bipartisan compromise, with an April vote in the Senate Committee on Banking, Housing, and Urban Affairs now possible. The stablecoin yield question is the single most consequential unresolved fight in U.S. financial regulation right now.

Every bank in America is deploying AI chatbots. A new industry report suggests most are deploying the wrong ones.

Glia's 2026 Banking AI Benchmarks Report, based on data from 400 financial institutions, found that purpose-built banking AI achieved accuracy rates above 92% in Glia's benchmark tests for interpreting financial terms without needing repetition — while a generic model might interpret "CD" as a compact disc rather than a Certificate of Deposit. That's not a quirky edge case; that's a customer calling about their savings being misunderstood by their own bank's chatbot. Banking-specific AI achieves a 94% containment rate in the report's tests — resolving routine tasks without escalating to a human, per Glia's report. Every escalated call costs a bank $8–$12. At millions of interactions, the ROI math is fast.

The real-world numbers back this up from both directions. Klarna reported its AI chatbot has handled over 2 million customer interactions, resolving refunds and payment-plan issues at speed. JPMorgan Chase announced a $1.2 billion AI spend this year — part of a nearly $19.8 billion technology budget — targeting fraud detection, customer service, and internal operations. The lesson isn't "AI bad." It's that a chatbot trained on the internet isn't the same as one trained on banking workflows, and the difference now has a dollar sign attached.

Private credit — the shadow banking system where funds like Apollo and Blackstone lend directly to mid-sized companies, bypassing traditional banks — has become a $3 trillion market. A huge chunk of those loans went to software companies, which looked like safe bets: predictable subscription revenue, sticky customers. That logic is cracking.

AI wiped $1 trillion off software stocks in February 2026, and now it's repricing the loans underneath them. UBS has warned that in an aggressive disruption scenario, default rates in U.S. private credit could climb to 13%. Cliffwater, a major private credit fund, is facing withdrawal requests exceeding 7% of assets; BlackRock, Blackstone, and Morgan Stanley have all limited or gated redemptions.

The irony is darkly precise: AI agents are eating the software subscriptions that were the collateral underpinning these loans. When an autonomous agent can replace a workflow that companies currently pay $50,000/year in SaaS fees for, the growth assumptions behind those valuations break. Now lenders are frantically deploying AI to figure out which of their AI-era loans are actually safe. The tool being used to navigate the reckoning is the same technology that caused it.

Imagine telling your phone "find and book me a flight to London for under $500 next month" — and having an AI assistant not just find the flight, but securely pay for it without you ever pulling out a card. Visa launched "Visa Agentic Ready" this week, a program to help banks test exactly this scenario. HSBC, Barclays, Santander, and Revolut are already on board for the European launch.

This isn't a thought experiment; it's a controlled test track for AI-driven commerce. Visa is using tokenization — swapping your real card number for a secure, one-time code — so the AI agent never handles your sensitive financial details. Separately, Visa is also experimenting with letting users pay on-chain gas fees (the small transaction costs on blockchains like Ethereum) from their normal bank accounts, removing a friction point that has kept mainstream users away from crypto-native apps. Both moves point the same direction: Visa is building payment rails for a world where software buys things on your behalf.

When you hear "crypto," you probably think of volatile coins. The biggest action this week was in something far more traditional: U.S. Treasury bonds. BlackRock's tokenized treasury fund, known as BUIDL, has surged past $2 billion in assets. What BlackRock is doing: taking ultra-safe government bonds, representing them as digital tokens on a blockchain, and selling them to institutional investors who want faster, 24/7 trading and settlement instead of waiting days for transactions to clear through old-school banking systems.

This isn't the only tokenization signal. Figure tokenized a $50 million pool of home equity loans, showing the concept is expanding beyond treasuries into mortgage and real-estate credit. And in Switzerland, St. Galler Kantonalbank tokenized its actual registered shares on the Ethereum mainnet — not a private ledger pilot, but real bank equity on a public blockchain under regulatory oversight. The quiet story of 2026 isn't crypto speculation. It's the core technology of crypto being absorbed into the plumbing of traditional finance, starting with the safest and most boring assets on the planet.

Mastercard Virtual C-Suite — Mastercard introduced an agentic AI suite designed to give small businesses executive-level financial intelligence, drawing on insights from 175 billion transactions processed on its network in 2025. Each agent acts as a digital CFO, CMO, or security officer. This is a product announcement, not a deployment report — but Mastercard's data moat makes it more credible than a startup pitch deck.

Brighty Banking API for AI Agents — European crypto-native platform Brighty launched developer-ready infrastructure that lets AI agents autonomously check balances, send payments, convert currencies, and reconcile transactions. This is the infrastructure gap the industry said couldn't be solved — an API that gives bots their own bank accounts. It's European and crypto-native, so the regulatory sandbox is different from the U.S., but the direction is real.

Tempo Mainnet — A blockchain called Tempo went live, positioning itself as payment plumbing for AI agents using stablecoins — optimized for tiny, high-frequency payments (think cents per API call). Coverage so far is mostly crypto blogs and online community discussions, so treat this as an interesting experiment, not proven infrastructure.

Norman AI Compliance Tool — Norman launched a generative AI solution that reads dense regulatory text and checks a firm's internal policies against it, surfacing gaps and remediation steps automatically — a practical example of how AI guardrails might land in compliance departments.

A buy-now-pay-later company filling out a bank charter application. A $3 trillion loan market being triaged by the same AI that broke it. A 150-year-old Swiss bank putting its shares on Ethereum like it's uploading a PDF.

Somewhere, a generic chatbot at a regional bank is confidently telling a retiree that their Certificate of Deposit is available at Best Buy.

Until next week. —The Lyceum

If someone you know is trying to make sense of where money and technology are colliding, forward this their way.

The FTC declared "unfair" AI an active enforcement target under Section 5 of the FTC Act — including AI in lending and financial services. Read → FTC Draws a Line: "Unfair" AI Is Now an Enforcement Target