Reuters prices AI cash sweeps at $79 billion a year in added US bank interest

The $79 billion is a spread, and the spread can be backed out. Paying 4% on $3.8 trillion costs $152 billion a year ^[13]. If the increase is $79 billion ^[2], banks are assumed to pay about $73 billion now ^[14], an average of roughly 1.9% on those balances ^[15]. Those are our back-calculations from the totals the Korea JoongAng Daily reports, and the implied starting rate is the figure to check against any bank's disclosed deposit costs.

The scenario moves the whole pool at once, so it is a ceiling. It is also narrower in count than the dollar figure suggests: 32 of the more than 500 banks Reuters covered ^[1] is at most 6.4% ^[16], and the 78-bank case is at most 15.6% ^[17], assuming both groups sit inside the 500-plus.

Consumers offer less delegation than the ceiling needs. In a Thales survey of 14,300 people in 13 countries, 34% would let AI cancel unused subscriptions and 32% would let it switch them to cheaper products ^[8]. Only 7% would let it transfer money between bank accounts ^[9]. Applying 7% to Reuters's $79 billion gives about $5.5 billion a year ^[19]. That is an illustration, since the survey counts people across 13 countries and not dollars of US savings.

The counter-case is that moving a balance may not need delegation at all. Meta's Muse, introduced last month, cancels services or switches users to cheaper alternatives with their approval ^[5]. If a balance transfer could be set up the same way, a tap would be enough and the 7% would understate the pool. Park Chan-am, who runs Stealien and advises the presidential council on national AI strategy, argues the other way. "The more authority an AI agent has, the greater the risk that those powers could be abused if the system is hacked," he said ^[10].

The Financial Times figure does not settle it. The paper estimated on Wednesday that as much as $500 billion across the US banking system could be at risk ^[18], and the JoongAng Daily's summary does not say whether that measures deposits or income, so it cannot be set beside Reuters's $79 billion.

Korea's MyData service shows the consent route working where a regulator designates the provider ^[6]. It reprices loans, so it shows consumers will accept automated rate-shopping; it does not show deposits moving. Korean savings and demand deposits stood at about 681 trillion won ($510 billion) at the end of July ^[12], and the service is expected to expand to debt restructuring and refinancing ^[11].

We think deposit stickiness is a live risk for bank investors and that $79 billion is the wrong number to model on today's permissions. We would not cut bank earnings by anything near that sum on this evidence; the exposure to size is each bank's share of savings balances paying well under 4%. We would be wrong if the share willing to let an agent move money climbs well above 7%, or if approval-based agents begin moving balances, since either would move the 32-bank figure from a ceiling toward a base case.