A Different Kind of Bank Run?
"Time is what we want most, but what, alas! we use worst..."
— William Penn
This week gave us another glimpse of our AI-powered future. AI personal agents are becoming a thing you need to understand, if not use.
Torsten Sløk of Apollo posed a question about the impact of these agents on banking. When everyone has a personal agent doing their shopping, will traditional banks lose deposits, as agents seek out higher interest rates?
Until recently, most of us used AI by asking it a question and waiting for an answer. But AI agents can act on the answers. Give one a goal, and instead of telling you what to do, it can do the work for you. Samantha, our VP of Operations, told me this week that she has “at least six” AI agents performing tasks for her in the background. I had no idea. She is finding ways to scale, boosting her productivity. She is an early adopter. Soon we’ll all have a personal AI.
Today there are still limits, of course. Broadly speaking, an AI agent can help you find a better interest rate. Compare insurance policies. Cancel the subscription you forgot about. Switch phone plans. Appeal a medical claim. Move your money.
None of these things is particularly difficult. But they take time.
Solving that annoyance is worth a lot of money. Businesses have a name for obstacles like these: friction. A surprising number of business models benefit from it. Banks hold cheap deposits. Gyms sell unused memberships. Subscription services renew long after they’ve been forgotten. And so on.
AI agents can remove much of that friction. Instead of reminding you that a better deal exists, they can increasingly go get that deal for you. And when the friction disappears, some very profitable customer behavior may disappear with it.
Agentic Bank Run
On September 8, Meta launched Muse, a personal AI agent. It quickly rose to the top of Apple’s App Store. Through a partnership with Plaid, it can access a user’s bank accounts, transactions, and investments.
Apollo’s chief economist, Torsten Sløk, immediately posed a consequential question. What happens when agents like Muse begin sweeping household cash out of checking accounts earning about 0.1% and into accounts paying roughly 3.3% to 5.0%?
His answer? An “agentic bank run”.
Opening a new account, linking it, moving direct deposit, updating automatic payments, keeping enough cash to pay bills… this is friction. Banks count on you not having the time or patience to do this. An AI agent can do it for you seamlessly, and as often as necessary to keep you earning a maximum yield.
The agent can stay within FDIC-insurance limits, sweep excess balances into higher-yield accounts, and move money back when it is needed. For one household, that may mean an extra few hundred or few thousand dollars a year.
Multiply that decision across millions of households, and banks could lose a large share of the cheap deposits they rely on to fund loans and investments. This is not a 1930s-style panic. It’s a different kind of bank run, driven by automated optimization.
If deposits leave in search of higher yields, it could force banks to pay a higher interest rate on deposits, squeezing margins and perhaps fueling more bank consolidation.
The higher yields are being offered by online banks (aka fintechs) but also by online brokers like Schwab and Fidelity. These entities are less encumbered by the overhead associated with bricks and mortar.
A loss of deposits could also expose weaknesses in bank balance sheets. Remember Silicon Valley Bank? Depositors withdrew enough capital that the bank was forced to sell some of its treasury bonds to cover withdrawals. The problem was those securities had been purchased during the near-zero interest rate period. As interest rates went up in 2022 and 2023, the value of the bank’s treasuries went down, and the bank suffered a $1.8 billion loss. That triggered a bank run.
Fast forward to today, with rates at 20-plus year highs. How many banks have paper losses on their treasury holdings? This question is largely irrelevant if the treasuries are to be held to maturity. But as we saw with SVB, when a bank needs to raise capital, it has to sell assets. If the sales result in actual losses, what we saw with SVB could happen across the country.
To Sløk’s point, what happens when a customer with $25,000 sitting in that bank earning almost nothing gets an AI agent that can move it somewhere earning 4% in a few minutes?
Nothing, probably. But what happens when a thousand customers get the same agent?
An AI agent doesn’t make the bank's bonds worth less. It makes the cheap, sticky deposits that allow the bank to patiently hold those bonds less dependable.
Banks could respond by paying depositors more to stay, of course. But when long-dated assets are earning roughly 3% and the deposits funding them currently cost 0.06%, that creates a different problem.
Banks are only one example of what happens when an AI agent makes it easier for customers to act in their own financial interest.
Who Lets Your Agent In?
On September 21, Amazon blocked Meta’s Muse from shopping on Amazon.com, raising privacy and security concerns around customer credentials and account information.[1] To be fair, those are legitimate concerns. If I give an AI agent access to my Amazon account and permission to spend my money, I want to know what it can see and what happens if it screws something up.
But that makes me wonder about everything else we just discussed. What happens if your bank decides it does not want an outside agent moving deposits? Or your insurer does not want one filing appeals? Or a hospital does not want one scrutinizing its bills?
Maybe they have legitimate security or privacy reasons. Maybe they have business reasons. Probably both. I suspect we are going to hear a lot more about this in the days ahead.
And there is one other question I would keep in mind: Who is the agent actually working for?
If it recommends a bank, insurance policy, or product because it is genuinely the best choice for me, great. If somebody paid the company behind my agent to put it there, I haven't gained an advocate. I've (once again) become the customer of big tech.
An AI agent with a fiduciary responsibility to you could become a powerful equalizer for consumers. Agents that can tackle friction will empower individuals, who often have day jobs and can’t take hours, days or weeks to do things like challenge insurance companies, shop for better interest rates, compare mortgages, or find the best prices and terms for major purchases like cars, appliances, etc. The potential opportunities are endless.
An agent doesn’t get tired of comparing rates. It doesn’t forget an appeal deadline. It doesn’t give up after 45 minutes on hold. And it doesn’t decide that saving $500 isn’t worth ruining its Saturday afternoon. But Meta doesn’t exactly have the best track record when it comes to societal good.
AI agents might give consumers a little more leverage. But only if they are truly working for you.
Are you using an AI agent? Let me know.
Thanks for reading,
Ed D’Agostino
Partner & COO
[1] https://www.retaildive.com/news/amazon-meta-muse-ai-agent-shopping-experience/831018/
Ed D’Agostino
Publisher & COO
If you prefer to listen to Global Macro Update, you can do so here:
Get Ahead of Big Market Trends—with Global Macro Update
Today’s most successful investors share one thing in common: They spot major trends early and move quickly to take advantage of them.
Twice a week, Ed D’Agostino—our publisher and COO—shares macro insights and analysis, along with exclusive interviews with leading economic, geopolitical, and investment experts to help you get ahead of major trends and market shifts.
By opting in you are also consenting to receive Mauldin Economics' marketing emails.
You can opt-out from these at any time. Privacy Policy
I am not using any of the AI agents. I am close to doing so though. I am just not confident enough yet to understand the boundaries surrounding the information location. Who can see my prompts, responses? What and where are my details stored, and does that differ from my already stored digital footprint? What digital supply chains are involved in processing my prompts? I see excellent use cases in generating training materials or business presentations, but at this stage, I feel too uninformed to dip my toe in the water. As for handing over my financials to an agent - that's an easy hard pass for me.
(side note - this is the first time I have commented. It would be interesting to see if you received more comments this week. I was pressed for time, so I am a week behind. Catching up now. Great thoughtful article. Thanks Ed.)
I've been running ChatGPT Pro for about two years now. I find it is excellent for many things, not so great for others. It is a huge time suck, for example. But as an editor it does extremely well. Most of the sugested rewrites are noted but generally only used as example corrections, even though it has claimed to "know my voice." Over the summer I loaded up a local model on my M4 Mac mini and began experimenting with it. For text interactaion it is pretty good but not great. I haven't tried tasking it with vibe coding or creating images, which I assume will bog down on the local machine, but that's in the works as I get more comfortable with the local mode. But clearly running local models, that keep your information local, is going to be the key. The data center AI model is basically the equivelent of the AOL/AT&T/IBM "information superhighway" hype of the 1990s. Yes, it was much more efficient compared to the Internet, but we were willing to accept a slower Internet because anyone could use it. The big datacenters are the new gatekeepers. And networks don't like gatekeepers.
Thanks Ed, an excellent and eye-opening article. Like you, I will not be using AI Agents, I wouldn't trust Facebook to tell me the time, or any of the other players for that matter. That isn't to say that I am unwilling to use AI. At 78, I have the misfortune to require the services of Cardiogists, but they are so Type A and rigid guideline thinkers. Getting them to listen rather than just read the test results ain't easy, and I am a retired physician!! Consequently, ChatGPT and I are building an App with Codex which presents, in a way that cannot be ignored, the associations of BP, pulse, arrhythmia, ECG, symptoms, events and wine intake! The capability of the AI is astonishing, not only in building the App, but in understanding the underlying medical circumstances and necessary algorithms. In the past few months, ChatGPT has resolved Tax & Pension issues, advised on a Project Closure Report, collated a 180 page historic document, and much more. However, I take nothing as complete until fully fact checked. Not sure how I will react if I am told to cut down the wine! ChatGPT has had remarkably few errors and hallucinations, but constant vigilance is necessary, and Agents are a step too far!
I am not using AI agents and I am in the software industry. They are too opaque, to unpredictable, make significant errors and fabricate information (which are papered over as 'hallucinations'), and are increasingly dangerous to humans and society. They are not safe for any kind of autonomous activity, and you can't even trust the results you get -- you have to fact check them. They are increasingly hard to control, signaling that we are behind the curve on making them safe for use.
A very useful commentary given the risks we are currently assessing in our own environment in South Africa where higher levels of corruption etc lend themselves to scams of all sorts which are difficult to manage without some basic knowledge. Although to be fair banking regulations here are of the best in the world.
I will never pay for AI. If I was starting a business, I'd absolutely use AI but I'd never get enough value to warrant monthly fees. I don't have unused subscriptions. I have my money in high yield accounts.
I do intend to use free AI to assist in investment evaluation but I have no desire to ever allow AI to make purchases on my behalf, access my financial data, etc.
Certainly plenty will but I'd be surprised if a high percentage of wealthy people give AI access to financial data without regulation to ensure their wealth is protected.
I would Never let an AI company have complete control of my BOA checking account to transfer $ to my Fidelity account. This is precisely what I want to do now as I have 21K sitting in my BOA Checking Account earning 0%. I don't mind spending 2-3 hrs figuring it out, including a visit to my BOA branch office. Once I know how to do it - i will write it down in my BOA & Fidelity folders, and it will become an effortless and SAFE task!
Bear in mind you need to feed to AI all your accounts info anyways - so how much time are you saving anyways???
Slok's comments appear very logical at first glance. However, I might suggest a deeper analysis into the demography of individual bank deposits. For example, I recently heard an officer for Bank of America say that its "average" consumer maintains a monthly average of $400 in his or her checking account, and these funds are used primarily for payments rather than investments. I'm guessing that the highest-income and highest-wealth cohorts already practice efficient management of transactional balances either individually or with the help of an advisor. I have no numbers to back up my views, and I haven't done the analysis. It's just a strong hunch. I strongly question whether we'll see an agentic run on banks, but it makes for a good headline.
I believe opening up your financial accounts to AI agents is a recipe for disaster as other rogue AI tools will target these areas and your assets will evaporate overnight. I am using Claude to streamline my investment analysis and only locally nothing in the cloud.
Thanks for the article Ed. Also have you considered another Camino - next spring we are doing the Via Francigena.
I have considered it, but it means telling the agent everything about the chunk of your life that you want it to manage. That raises a bunch of fiduciary and privacy concerns. In the case of Meta, I would NEVER let them into my life - social media of otherwise. I asked Claude (which already knows a great deal about me), and its answer was telling. Don't do it!