Apollo’s Torsten Slok warns AI cash-sweeps could erode banks’ cheap deposits
He points to a 0.1% checking average versus 3.3%–5% alternatives as stablecoin rails like x402 enable agent payments.
Apollo chief economist Torsten Slok warned that AI assistants could automate household cash out of low-yield checking and into higher-yield alternatives, creating a “slow-motion bank run.” The same agentic automation theme is already live on stablecoin payment rails, where Coinbase’s x402 is described as enabling agent-driven payments without cards or human approval.
Slok’s “Slow-Motion Bank Run” Thesis: AI Turns Rate-Chasing Into Autopilot
Slok’s warning is about plumbing, not panic. Banks fund loans with cheap deposits, and checking balances are the cheapest part of that stack.
In a Sunday note titled “Is an Agentic bank run coming?” Slok argued that agentic AI assistants could turn deposit shopping into continuous automation. He framed it as a “slow-motion bank run” where money leaves not because depositors are scared, but because software is paid to optimize.
The incentive is clean math. Slok cited a 0.1% national average on checking accounts and contrasted it with deposit products paying roughly 3.3% to 5% annual interest, naming Revolut, SoFi, Varo, LendingClub, and Wealthfront. On his $10,000 example, that spread is the difference between about $10 a year in checking and about $100 a year in higher-yield alternatives.
That delta is small enough to ignore manually and large enough to automate. Agentic finance, as Slok described it, is AI that acts rather than just answers: monitoring balances in real time, comparing returns across institutions, sweeping idle cash into higher-yield accounts, then moving it back in time for bills. The second-order effect is the point. If households outsource this behavior at scale, banks lose the low-cost funding they rely on to make loans.
Slok put the systemic-risk claim plainly: “If every household used AI agents to optimize the return on their cash balances, banks could lose a large share of the cheap deposits they rely on to make loans, which would be a problem for the entire financial system,” he wrote. He referenced Meta’s personal agent “Muse” as an example of the kind of assistant that could soon make this behavior default.
Market sizing around agentic AI is still noisy, which matters because adoption speed is the whole trade. Mordor Intelligence estimated agentic AI in financial services at $7.78 billion in 2026 and projected $43.52 billion by 2031, while MarketsandMarkets pegged the narrower “AI agents” segment at about $845 million in 2025.
Crypto Rails Are Already Agent-Ready: x402’s Stablecoin Payment Stack and Reported Usage
The crypto linkage here is not theoretical. Stablecoin rails already support agentic payments, which is the other half of the automation loop: not just deciding, but executing.
Coinbase’s x402 protocol was described as an agentic payment standard that lets an AI agent pay for online services in stablecoins within seconds, with no account, card, or human approval. That matters because it removes the friction points that usually slow automated money movement in legacy payment flows.
Reported usage suggests this is being productionized, even if the exact scale is fuzzy. The x402 protocol has reportedly processed approximately 188 million to over 205 million cumulative transactions and has around 69,000 active agents, with no single definitive figure or timestamp provided alongside the range.
Institutional participation is also moving toward standard-setting. Cloudflare, Google, Visa, Mastercard, AWS, Circle, and Stripe have joined the x402 Foundation, which the Linux Foundation now governs. The membership list is the tell. If major payments and cloud firms are showing up in the same room under a governance umbrella, the goal is interoperability and distribution, not a demo.
How I’d Trade the Narrative: Bank-Funding Stress Meets Stablecoin Payment Adoption
The threshold that matters is whether agentic cash management becomes a default consumer feature, not a niche optimization. If checking stays anchored near ~0.1% while alternatives hold ~3.3%–5%, the incentive to sweep does not need a bull market to work. It just needs permissioning and habit.
The real test is measurement. If x402 (or any comparable rail) starts publishing timestamped, narrowing disclosures that confirm growth beyond the current ~188M to >205M transaction range, and if consumer assistants like Muse ship cash-sweep functionality, the setup starts to look structural rather than narrative-driven. In practical terms, this matters when automated deposit betas rise and stablecoin payment standards show verifiable, accelerating throughput at the same time.
