AI’s Real Job in Consumer Banking Expereince: Retiring Resolution Debt

For twenty years, consumer banking’s ambition has been to make banking simpler, faster and more digital. We moved transactions from branches to apps, removed clicks and digitized forms, and we got very good at it. But the interface was never the hard part. Underneath it sits a division of labor that has barely changed since the branch era: the customer is responsible for knowing what she needs, knowing the bank offers it, and asking at the right moment. The bank waits. 

That gap deserves a name. Call it resolution debt: the distance between what a customer needs resolved and what the institution resolves without being asked. It is not a usability problem. It is unpaid work, the monitoring, the diagnosis, the product selection and the timing that the bank is better positioned to perform and has quietly assigned to the customer instead. Like technical debt, it compounds, and someone eventually services it. Unlike technical debt, it does not sit on the bank’s balance sheet. It sits on the customer’s sheet. 

Twenty years of digitization paid down transaction cost and left resolution debt untouched. That is why banking still feels like work after we digitized nearly all of it. We were servicing the wrong ledger. 

Banks design journeys: open an account, replace a card, dispute a transaction. Customers think in outcomes—”I need to make sure I have enough money until payday.” The gap between the two is where the debt originates. We require the customer to translate her intent into our products and processes, and we charge her the difference in effort, in money, or in both. AI creates the possibility of reversing the equation: the bank interprets the intent and assembles the experience around it. 

Where the Debt Actually Sits 

That is not a service-desk problem. Resolution debt shows up in four places, each with a P&L line attached. The shortfall the bank could see coming and does not flag, which becomes an overdraft. The credit decision it could make from its own deposit data but waits for an application and then declines on a bureau file. The balance sitting at a rate the customer would move if she were paying attention. The dispute she has to detect, evidence and chase, because detection was left to her. 

None of that comes from a shortage of AI. It comes from structure: her money sits in a deposit ledger, a card ledger and a servicing platform that hold no shared view of her, so she becomes the integration layer, carrying context between systems that will not speak to each other. 

The Uncomfortable Part: The Debt Pays Us 

Structure explains how the debt accumulated. It does not explain why an industry this capable has left it in place for two decades. The honest answer is that a meaningful share of consumer banking revenue is collected at precisely the moment resolution fails. 

Overdraft is the clearest case. Even after policy changes cut them by more than half, consumers paid roughly $5.8 billion in overdraft and NSF fees in 2023, according to the Consumer Financial Protection Bureau. Most of those fees were charged for events the bank’s own systems could have predicted days earlier. We are not merely failing to warn her. We are billing for the silence. 

Deposit pricing works the same way, more politely. A bank with a low deposit beta passes almost none of a rate rise through to savers and can do so only because customers do not move. Curinos estimates that as much as a third of shareholder value has historically depended on savings balances held in place, largely by inertia. 

Servicing metrics complete the picture. Containment counts an interaction as a win for staying in self-service, handle time for ending quickly, and neither asks whether anything was resolved—so effort transferred to the customer is booked as efficiency gained by the bank. 

This is the real obstacle, and it is not technical. Any AI that genuinely retires resolution debt will cannibalize fee income and raise the cost of deposits the bank now gets cheaply. That is a strategy conversation, not an engineering one—and it is why so many programs stop at a better front end. A chatbot is safe precisely because it changes nothing about who bears the work. 

The bet on inertia is also weakening. Cerulli projects $124 trillion will change hands through 2048, most of it to heirs who have never carried this debt anywhere else, and Curinos argues inertia is already eroding as agentic AI begins shopping on customers’ behalf. When the customer’s agent reads her rate and moves her balance, rate-insensitive funding stops being a strategy. 

Why AI Can Repay It and Digitization Could Not 

Banks have wanted to close this gap for years. What changed is that reasoning across fragmented systems no longer requires consolidating them, so a context layer can sit over the deposit, card and servicing platforms rather than wait on a decade-long migration. Credit shows what that is worth. The Federal Reserve’s work on alternative data describes “invisible primes“—borrowers whose bureau files understate them but whose cash flows do not. The deposit account already holds that data. A bank that reads it can offer terms before an application exists, to a customer it would otherwise decline. 

Refinancing Is Not Repayment 

This is where most AI programs go wrong. A conversational front end lowers the interest rate on resolution debt—she states her problem in her own words instead of hunting through a menu—but she is still servicing the loan. The Consumer Financial Protection Bureau has found that chatbots handle basic inquiries but falter when customers face complexity or cannot reach a human. Automation without resolution is not customer experience. It is refinancing. The discipline runs the other way too: the best AI interaction may sometimes be the one that sells nothing. 

The ceiling is real: recommending the wrong movie is irritating; misstating a balance or a credit decision is not. When the OCC, Federal Reserve and FDIC updated interagency model-risk guidance in April 2026, they placed generative and agentic AI outside its scope while reinforcing that governance must be commensurate with risk. 

Retiring principal takes three decisions, and only one of them is technical: 

• A context layer that earns its keep. One resolved view of a customer’s balances, obligations and cash flow that any channel can read—and that underwrites and prices before she thinks to ask. 

• Intent-to-resolution as the service metric. Retire containment and handle time. Measure the elapsed effort from the moment an intent forms to the moment it is resolved, counting every restatement of context as a penalty—then publish it and pay people against reducing it. 

• A revenue plan that survives resolution. Decide explicitly what replaces the fee income and cheap funding resolution debt now produces. Programs that skip this get quietly defunded when the trade-off surfaces. 

The constraint is almost never the model. It is the willingness to give up revenue that depends on the customer not knowing something. 

What Carrying the Debt Costs 

The bill for inaction arrives on the funding side. Retail balances are the base that lets a bank price corporate credit competitively and lean less on wholesale markets. If that base erodes materially, funding costs rise and loan pricing gets less competitive. A retail experience failure becomes an institutional pricing failure, one quarter at a time—and it will be reported as a margin story rather than a customer story, which is precisely why it will be diagnosed late. 

The default is not displacement but fragmentation. Customers do not leave a bank carrying high resolution debt; they route around it, until the bank holds the account with the lowest balance and the highest cost to service, and calls it retention. 

Back to the Customer Who Started This 

Return to the customer trying to make it to payday. She opens the app, scrolls transactions, tallies what is still due and guesses—and if she guesses wrong, the bank charges her for it. The institution may already hold what she needs and still waits to be asked. It should recognize the shortfall before it lands, explain what changed, lay out the options, ask permission and act. 

She is the whole argument. Most of what is needed to help her already exists inside the bank; what is missing is the decision to use it on her behalf rather than at her expense. That makes this a business-model question, not a technology one. Every institution should be able to state how much of its consumer revenue depends on customers not knowing something. Almost none can. 

The question is not whether banking becomes intelligent. It is whether that happens inside the institutions holding the deposits today.