Apollo chief economist Torsten Slok warned about the possibility of an “agentic bank run”: AI agents continuously searching for better returns and automatically moving consumers’ deposits from low-yielding bank accounts to higher-yielding alternatives.
The phrase is deliberately provocative, but the underlying issue is whether AI materially reduces the inertia that has historically made deposits sticky.
My take: The money movement part is real (it's been happening for 10+ years without AI agents). The optimization part is hype—constant account shifts carry costs, and the gains are smaller than advertised. Better financial outcomes require people to change behaviors, not just accounts. Agents might optimize rates and fees, but they'll be doing it against institutions running their own agents—agents that have run the same negotiation a million times and know exactly how much they need to concede. Whether consumers come out ahead is going to be a hard question to answer.
The real contest isn’t just one agent against another over a single customer’s savings rate. It’s a battle for which side can scale to millions of negotiations the fastest. Meta and Google will be doing it across their entire user base. A $2 billion bank would be doing it with whatever its core provider ships, on the core provider’s timeline.
AI agents will move money. And why shouldn’t they? People have been moving money for years.
Checking accounts have become paycheck motels. Money lands, sits for a few days, and moves on. $3 trillion was displaced from banks and credit unions to Robinhood, Fidelity, SoFi and their fintech brethren with no AI agent involved. Autopay and card-on-file updates are real friction, but they’re speed bumps.
Personal AI agents will take care of the easy stuff. Opening new accounts? Piece of cake. Transferring balances? Child’s play. Figuring out who’s got better rates or lower fees? Easy-peasy. Determining who has the better account? Not so fast, bucko.
Optimization is a technical term and requires something called an objective function—a number to maximize. Consumer finance doesn’t have one, and its pieces conflict:
Two big challenges here:
An investment portfolio can be graded. There’s an index, a peer group, a risk-adjusted return, and a methodology for deciding whether the manager earned the fee.
A consumer household has none of that. You can’t run your financial life twice. There’s no version of the last three years where you kept the other card, stayed at the other bank, and left the money alone, sitting next to this one for comparison. There’s no base case, so the agent’s performance can’t be objectively verified. Nobody can prove it added (or subtracted) value.
Hard to defend for a product that promises optimization. Easy to live with for a product that wants to keep charging for it.
Say the objective ends up being yield, which is the likeliest default. Follow it where it goes.
Agents didn't create the fragmentation. A millennial couple like Cameron and MacKenzie already has ~40 financial relationships, and no AI agent helped.
What an agent changes is the rate of turnover. Rates move constantly and they don’t move together, so an account that’s best in March is fourth-best by June. An agent chasing the leaderboard doesn’t settle at 40 accounts. It runs 40 at any given moment and a hundred over the course of a year, opening and abandoning as the rankings shift.
Chasing the top rate beats sitting still, but the question is by how much.
Getting to real money takes real balances. At $25,000, a 100-basis point difference is $250 pre-tax, which is a nice bump. But a household capturing a full point was parked at 0.01% to start with, and fixing that is one move, not a hundred. Once the money is sitting at a competitive rate, the spread between best and decent runs 30 to 50 basis points—$75 to $125 on that same $25,000, and less after tax.
Top rates are also advertising. The leader is usually buying deposits with a promotional rate that steps down after four to six months, so the agent captures the differential for part of the year and then moves again.
The gain is real but small for most households. The costs come with every opening:
The agent has to optimize something, so it optimizes what carries a number: rate, fee, cash back percentage and minimum balance. These are all published, structured and comparable. What gets left out of the calculation are the intangibles:
None of these can be evaluated from the outside. You find out after you're a customer.
The objection is that an AI agent can read Yelp, Reddit and Trustpilot. Those are other people’s claims under other people’s circumstances, public because someone was angry enough (or incentivized) to post.
If personal AI agents treat reviews as an input to money movement, manipulating—or even falsifying—reviews skyrockets. The technology that makes the agent possible is the same technology that makes a convincing fake review easy to produce.
Agents will be good at the measurable part, and they'll report improvement. Consumers looking for optimization won't know the difference.
Agents might not be that good at the measurable part.
Every AI product on the market ships with a disclaimer that its output can be incorrect. I saw one when I asked Claude to look over this post and tell me what readers would argue with.
The disclaimer means the user has to do the verification. Checking the agent's work means doing the agent's work. A human capable of checking the recommendation doesn’t need the agent. And a human who can’t check it has no way to grade it.
When every consumer has a personal AI agent, every institution will have one, too—one that reprices continuously. That creates two problems for the personal agent:
Two machines at the table and one of them has the file.
Financial institutions will also be able to see who has an agent. Agents leave tracks: constant API queries against the account, retention offers answered in seconds instead of days, balances moving in round numbers on a schedule. A consumer who generates those signals is telling the bank they’re ready to walk and gets priced accordingly. A consumer who generates none of them is telling the bank they aren’t going anywhere and gets priced accordingly too.
Handing agents to everyone doesn’t level the field. It sorts customers into the ones who’ll leave and the ones who won’t, and the second group pays for the first.
Insurance already ran this experiment. Carriers built models to identify which customers were least likely to shop and charged those customers more. Regulators called it price optimization and killed it.
Deposit and fee pricing hasn’t faced that test—yet. An institution that quotes 4.3% to a household with an agent and 3.1% to an identical household without one is doing what insurers were told to stop doing. Elizabeth Warren will be all over this.
The real contest isn’t just one agent against another over a single customer’s savings rate. It’s a battle for which side can scale to millions of negotiations the fastest. Meta and Google will be doing it across their entire user bases. A $2 billion bank would be doing it with whatever its core provider ships, on the core provider’s timeline.
The profitability of a lot of financial products depends on customers leaving value on the table:
If every customer claims everything, those products don’t survive at today’s prices. They get repriced, and the arbitrage closes.
Mark Zuckerberg thinks consumers with personal AI agents will, in the aggregate, come out ahead. They won’t. If products get repriced until the arbitrage closes, the gains don’t accrue—they get redistributed.
The customer who extracts $1,200 from a $695 card loses when it gets repriced. The net effect is a transfer from sophisticated customers to unsophisticated ones, plus whatever product variety disappears when the cross-subsidy stops working.
Agents are going to manage money. Measuring how well they do it is a problem yet to be solved, and I’d want it solved before I hand over the keys to my accounts to a personal AI agent. OK, before my wife hands over the keys.