Agentic AI could change more than how businesses operate. It may change how customers evaluate relationships, make financial decisions, and determine which companies deserve their business.
For decades, banks have benefited from something that rarely appears in a strategy deck: customer inertia.
People do not continuously optimize their financial lives. They leave excess cash in low-yield accounts, tolerate fees they could avoid, and stay with institutions when better offers exist elsewhere. Some of that is loyalty. Some is trust and the value of an established relationship. But some of it is simply that customers have other things to do, and the effort required to evaluate alternatives often outweighs the perceived benefit.
In September, Apollo Chief Economist Torsten Sløk raised the prospect of an "agentic bank run," in which AI assistants automatically move household cash toward higher-yielding alternatives. He posited banks could lose substantial amounts of low-cost deposits if millions of consumers began using agents to optimize their cash balances. It's a provocative scenario, particularly given what we've already witnessed when digital banking makes it possible to move money quickly.
But I think the discussion risks focusing on the wrong consequence. AI agents don't need to cause a bank run to change banking economics. They only need to make customer inertia less dependable, and the implications extend well beyond financial services.
When the customer no longer has to do the work
Imagine an AI agent operating on behalf of a household. It understands income, recurring expenses, liquidity requirements, savings objectives, outstanding debt, and financial preferences. It can evaluate alternatives continuously, without requiring the customer to initiate every comparison or spend time researching available options.
Someone with $25,000 sitting in a checking account might occasionally wonder whether some of that money should be earning a better return. An agent could evaluate that question daily, considering interest rates, liquidity needs, deposit insurance, upcoming expenses, and the value of the customer's existing banking relationship. It might recommend moving $15,000 into a higher-yielding account, or it might conclude that keeping the money where it is makes more sense. Either way, the customer no longer has to perform the analysis.
I've spent much of my career leading technology organizations in financial services, including periods when digital banking fundamentally changed how customers interacted with institutions. We generally evaluated those investments by asking what technology allowed the bank to do better: reduce operating costs, improve service, accelerate transactions, or reach more customers. Those were important objectives, but some of the more consequential changes came from what technology allowed customers to stop doing.
Mobile banking eliminated the need to visit a branch for many routine transactions. Digital account opening reduced the effort required to establish a new relationship. Online lending made it easier to compare offers and apply for credit. Agentic AI could remove another constraint: the need for customers to keep paying attention to opportunities to improve their financial position.
That matters because attention has economic value. Businesses have long operated in environments where customers don't have the time, information, or motivation to evaluate every available alternative. Customer retention often reflects a combination of genuine satisfaction, established relationships, convenience, and the simple fact that changing providers requires effort. Agentic AI could begin to alter that balance.
The economic implications extend far beyond banking
Consider the automobile market. A consumer walks into a dealership considering a new vehicle with a monthly payment exceeding $1,000. Today, that decision may be influenced by advertising, financing promotions, trade-in offers, and the customer's immediate desire to purchase. The dealer and lender naturally focus on completing a transaction, while the consumer may focus primarily on whether the monthly payment fits within the household budget.
An agent acting in the customer's financial interest might approach the decision differently. It could evaluate total cost of ownership, depreciation, insurance, interest expense, the condition of the existing vehicle, and the household's longer-term financial obligations. It might recommend a less expensive vehicle or identify better financing terms. It might also conclude that replacing the current vehicle is a poor financial decision, even if the customer can afford the monthly payment.
That last possibility is important. We're accustomed to discussing AI as a way to help consumers find better products or prices. But an agent genuinely working for the customer might recommend avoiding the transaction altogether. The economic impact would extend beyond which lender receives the loan or which dealership sells the vehicle.
Insurance presents a similar opportunity. Most consumers don't routinely compare their policies against competing offers, especially when the savings seem modest, or the coverage differences are hard to understand. An agent could evaluate premiums, deductibles, exclusions, coverage limits, claims history, and provider reliability. It might identify a better policy, negotiate a renewal where possible, or conclude that the current provider offers sufficient value to justify staying. The insurer would be competing against a more informed customer, without another insurer necessarily being directly involved in the conversation.
Subscription businesses offer an especially useful example because we're already seeing the early stages of this change.
Applications such as Rocket Money, PocketGuard, and Hiatus help consumers identify and monitor recurring charges, understand subscription spending, and recognize services they may no longer need. Rocket Money can detect subscriptions from linked financial accounts and, through its paid service, offer cancellation assistance for eligible subscriptions. PocketGuard identifies recurring bills and subscriptions, while Hiatus provides tools to monitor subscriptions and spot changes in spending.
These applications address a problem subscription businesses have benefited from for years. Consumers accumulate streaming services, software licenses, gym memberships, digital subscriptions, and other recurring expenses, often without maintaining a clear picture of their total cost. A $12 or $20 monthly charge may not warrant much attention individually, but those charges accumulate. Customers may keep paying long after the original need has disappeared.
Today's subscription-management applications largely help customers recognize and manage those expenses. Agentic AI could extend that capability by evaluating whether each subscription still provides sufficient value, comparing alternatives, identifying overlapping services, and potentially initiating cancellation or renegotiation within the authority the customer grants.
For a subscription business, the consequence is straightforward. Retention may depend less on customers forgetting to cancel and more on whether the service continues to justify its cost. That isn't necessarily a bad outcome for consumers, but it could challenge business models that assume a certain amount of passive renewal.
Across banking, insurance, automotive, retail, and subscription services, the underlying economic question is similar. How much of what companies currently recognize as customer lifetime value depends on customers actively choosing to maintain a relationship, and how much depends on the effort required to reconsider it?
I don't think most organizations can answer that confidently.
What happens when the customer has an intelligent intermediary?
Most enterprise discussions about agentic AI begin inside the organization. Executives want to know how agents can improve employee productivity, automate workflows, accelerate customer service, or reduce operating expenses. Those are legitimate opportunities, and many will create substantial value. But they reflect an enterprise-centered view of AI adoption.
Customer-side agents reverse the direction. A bank may deploy agents to improve deposit retention while customers use agents to identify better places to keep their money. An insurer may use AI to optimize pricing while policyholders use it to evaluate competing coverage. A retailer may optimize recommendations and conversion rates, while consumers use agents to decide whether to make a purchase at all.
Both sides become more sophisticated, but they aren't necessarily optimizing for the same outcome.
This creates a strategic challenge that goes beyond conventional digital competition. For years, companies have invested in reducing friction within their own customer journeys. They made purchasing easier, payments faster, account opening simpler, and renewals more convenient, generally expecting these improvements to increase conversion and retention.
Customer-side agents could reduce the friction involved in leaving, switching, negotiating, or deciding not to transact. The same technologies that make it easier for a company to serve customers may make it easier for those customers to challenge the company's value proposition.
The competitive question may eventually become less about how effectively an enterprise deploys AI internally and more about how it competes when customers use AI to evaluate the enterprise independently.
The harder problem is authority
Another issue I believe will prove considerably more difficult.
Recommending a financial decision is one thing. Executing it is another. An agent can tell a customer that refinancing a loan might save money, but what happens when the agent is authorized to initiate the refinancing process, provide financial information, compare final terms, and execute agreements? What exactly did the customer authorize, and how does the receiving institution establish that the authority is legitimate?
Consider an agent managing household liquidity. Did the customer authorize it to move $500, $5,000, or $50,000? Can it open an account, cancel an insurance policy, negotiate a contract, or commit the customer to a recurring obligation? What happens when the agent encounters circumstances the customer didn't anticipate when granting permission?
These aren't simply authentication questions. Knowing an agent's identity doesn't establish the scope of its authority, and establishing authority doesn't guarantee the quality of the decisions it makes.
In banking, we've spent decades developing controls around human and institutional authority: mandates, entitlements, transaction limits, approvals, exceptions, and escalation procedures. Those controls were built around people, organizations, and established transaction models. They weren't designed for autonomous software that continuously evaluates opportunities and may act across multiple institutions.
Now extend that challenge into the broader economy. A consumer's agent may interact with a bank's agent, an insurance provider's agent, or a retailer's agent. Each operates under different incentives, permissions, and responsibilities. Institutions will need ways to establish which agent is acting, whose interests it represents, what actions it is authorized to perform, and how that authority can be limited, revoked, or audited.
There's also a distinction between an authorized action and a sound decision. An agent might operate entirely within its permissions and still make a poor judgment, misunderstand a customer's priorities, or create an unintended financial consequence. Responsibility for those outcomes becomes difficult to allocate when multiple agents, service providers, and institutions participate in the same transaction.
We have considerable experience governing systems that execute predefined instructions. Governing systems that exercise delegated judgment presents a different problem, particularly when their decisions cross organizational boundaries. I don't believe we've resolved it, and I would be skeptical of anyone suggesting otherwise.
Whose economics should the agent optimize?
A tension underlies all of this that businesses will eventually have to confront: what happens when the best decision for the customer is not the best economic decision for the company serving that customer?
Consider a bank that provides an AI-powered financial assistant. The assistant recognizes that a customer would be financially better off moving a substantial deposit to another institution. Should it recommend the transfer? What if the same institution has invested heavily in AI to improve deposit retention?
Or imagine an automotive lender whose agent determines that a prospective borrower would be better served by keeping an existing vehicle rather than taking on another five years of payments. The recommendation may be financially sound for the customer, but it doesn't produce a loan for the lender.
These conflicts aren't new. Financial advisors, lenders, insurers, and retailers already operate in environments where customer interests and institutional incentives don't always align. Agentic AI could make those conflicts more visible and more frequent, especially as customers begin relying on software to evaluate decisions they previously made with limited information.
It also raises questions about the emerging business models for consumer agents themselves. Who pays for the agent? Who controls its objectives? Are recommendations influenced by referral fees, commercial partnerships, or the economic interests of the company operating it? Calling something a personal AI assistant doesn't establish that it acts exclusively in the customer's interest.
As these systems move from providing recommendations to exercising authority, understanding whose interests they serve will become as important as understanding what they're technically capable of doing.
Competing when inertia becomes less dependable
I don't believe the answer is to make it harder for customers or their authorized agents to change providers. Artificial friction may delay decisions, but it's difficult to imagine it as a durable competitive advantage if technology continues reducing the cost of comparison and switching.
Nor do I believe every customer relationship will become a continuous auction decided solely on price. Trust matters. Service matters: convenience, reliability, security, and the value of an integrated relationship matter. An intelligent agent may reinforce an existing relationship after evaluating alternatives, particularly when the relationship offers advantages that aren't apparent from a simple price comparison.
A bank that understands a customer's deposits, credit, investments, and liquidity needs may offer more value than an institution advertising a slightly higher savings rate. An insurer with dependable claims service may justify a higher premium. A subscription service that customers consistently use and value may have little reason to fear more frequent evaluation.
But those relationships may increasingly have to withstand scrutiny rather than rely on habit.
For executives, this introduces an uncomfortable strategic question: how much of our current economics depends on the customer making an affirmative decision to stay, and how much depends on the customer never seriously reconsidering the relationship?
It's a question worth examining before customer-side agents become commonplace, because the answer may be more consequential than many of the productivity improvements currently dominating enterprise AI investment discussions.
The change worth watching
The prospect of an agentic bank run makes for an attention-grabbing headline. It describes a potentially serious risk, particularly for institutions whose funding depends on deposits that could become more sensitive to changing conditions. But the broader economic shift may occur without anything resembling a bank run.
It may emerge through millions of ordinary decisions: a loan not taken, an insurance policy replaced, a subscription canceled, a purchase postponed, or a financial relationship retained because it genuinely offers the best overall value. Each decision may appear insignificant in isolation, but collectively they could change how businesses acquire customers, price products, manage retention, and measure relationship value.
I'm not convinced the transition will happen quickly or evenly. Consumers will differ in how much authority they're willing to delegate, and institutions have legitimate concerns about security, fraud, liability, and regulatory obligations. Some decisions will remain deeply personal, no matter how capable the technology becomes. Others may prove easier to delegate than we currently expect.
But I believe we're examining this change too narrowly when we focus primarily on how enterprises will use AI to improve their own operations. We should be spending at least as much time considering how AI will change the behavior of the people those enterprises serve.
Customers don't need to become financial experts or spend more time comparing alternatives. Software can increasingly perform that work and eventually may be authorized to act on some of its conclusions.
When that happens, businesses may discover that some of what they considered loyalty was actually friction. Others may discover their relationships are more durable than expected. We probably won't know which is which until customers have agents working for them.
