AI agents chasing 5% yields could drain banks' 0.1% checking deposits, Apollo's Slok warns

Slok's case begins with a yield table. A $10,000 checking balance at the 0.1% national average earns roughly $10 a year, and the same money in a 5% account earns $500, according to his figures as reported by Cryptopolitan ^[2]. The difference is $490 a year on every $10,000, a 50-fold gap ^[1][2]. The worst account on his list, at 3.3%, still pays $330 ^[3][3].

Banks have been able to leave checking near zero because the retail model assumes most of the money never moves ^[4]. In his September 27 note, Slok argues that an agent removes that assumption ^[6]. A household that leaves $10,000 idle at a $490 annual cost has made a choice (or rather, has declined to make one), and software told to optimize cash has no reason to wait. The danger in what he calls an "agentic bank run" is timing: if millions of people let agents chase rates, nothing stops all of those agents from acting in the same week ^[1].

Slok is applying to checking accounts a worry regulators have already raised about markets. Gary Gensler, then SEC chair, warned in a 2023 speech that a few dominant AI base models could make "similar decisions because they are getting the same signal" ^[10]. The Bank of England, in its April 2025 Financial Stability in Focus paper, wrote of AI strategies pushing firms into "increasingly correlated positions and acting in a similar way during a stress, thereby amplifying shocks" ^[9].

Against the Fed's new 3.75% to 4% range ^[8], national-average checking sits 3.65 to 3.9 points below the policy rate ^[4]. That spread is the cheap base Slok says banks would lose ^[7]. Adelfi's 5% is a full point above the top of the Fed's range ^[3][5]. The note, as reported, does not say whether the top rates cap balances or how many dollars sit at the national average.

Muse, the agent Slok cites, can read accounts but cannot yet send money ^[5], so for now the exposure waits on a single capability. I see three ways it resolves. Agents gain transfer rights, users switch them on, and cheap deposits leave in a cluster; this is the run Slok describes. Banks could instead see the permission coming and lift checking rates to hold balances. Nothing runs in that case, but funding still costs up to $490 a year more per $10,000 at the top of his list ^[1]. Or users never hand an agent authority over their savings, and the gap stays where it is ^[2].

I think the second outcome is the likeliest and the costliest for bank earnings, since a margin loss needs no panic, only agents quoting better rates. The counter-case is that a run moves faster than a rate change: if transfer-capable agents arrive on several platforms at once, banks will not get the time to reprice. The thesis fails if agents that can move money go live and checking balances hold still.