How AI-powered financial risk management...

Banks have always been in the business of managing risk. Every loan they approve comes with the possibility that the borrower may struggle to repay it. Every commercial relationship carries some degree of uncertainty, and even the most financially stable customer can experience a sudden change in circumstances. Banks have spent decades developing systems to measure these risks, from credit scoring models to portfolio monitoring and stress testing. The problem is that much of this intelligence still relies on information that tells the bank what has already happened.

A borrower misses a payment, their credit score deteriorates, or their account is flagged because a predefined threshold has been crossed. By the time that information reaches the appropriate team, the customer's financial position may have been deteriorating for months. The bank can respond by contacting the borrower, restructuring the loan or initiating collections, but the opportunity to intervene earlier may already have passed.

Artificial intelligence could change how this process works. Banks already collect information through transactions, loan servicing systems, account activity and customer interactions. With appropriate permissions and controls, AI can help connect relevant signals and identify patterns that suggest financial pressure is developing. That could give banks an opportunity to investigate emerging risks earlier, understand which customers may need assistance and make decisions while more options remain available.

The potential becomes clearer when we look at the scale of the problem. According to the Federal Reserve Bank of New York's August 2026 report, American household debt stood at approximately $18.77 trillion in the second quarter of 2026. Of that, credit card balances accounted for $1.26 trillion, while auto loans accounted for $1.71 trillion. The report also found that 4.7% of outstanding household debt was in some stage of delinquency.

Those numbers explain why early warning systems deserve more attention. Even a modest improvement in how banks identify and manage emerging financial distress could have meaningful consequences for lenders and borrowers.

Banks Have More Information About Financial Stress Than They Can Currently Use

A bank may have access to a remarkable amount of information about a customer's financial activity. Depending on the relationship and applicable permissions, it may be able to observe salary deposits, recurring expenses, loan repayments, account balances and changes in transaction patterns. For business customers, it may also have access to financial statements, cash-flow information and other records relevant to lending decisions.

The difficulty is that these signals are often distributed across different systems. Transaction monitoring may sit within one department, loan servicing within another and customer relationship management somewhere else. Each system is designed to perform its own function, and information may reach the risk management team only after a particular event has occurred.

Consider a customer who has maintained a healthy repayment history for several years. Their salary continues to arrive, but their average monthly account balance begins declining. Credit card utilisation increases, and they start relying more heavily on short-term borrowing to cover ordinary expenses. They may still be making every payment on time, so a traditional delinquency-based monitoring system has little reason to raise an alarm.

An AI system could potentially recognise the combined pattern as a reason for further assessment. Of course, these changes do not necessarily mean the customer is heading towards default. A large purchase, a temporary expense or a deliberate financial decision could produce similar signals. The purpose of the model would be to identify accounts that deserve attention, while the bank determines whether the pattern actually indicates financial difficulty.

This distinction is important. During our mentorship under Phaneesh Murthy, one of the recurring themes in our discussions is how organisations can use the information they already possess more effectively. Banks have invested heavily in digitisation, but information becomes useful only when it reaches the right decision-maker with enough context to support an appropriate response.

The Numbers Show Why Waiting for Delinquency Is Expensive

The Federal Reserve Bank of New York's second-quarter 2026 data provides a useful picture of the challenges lenders face. Credit card balances increased by $21 billion during the quarter, while auto loan balances increased by $28 billion. The report also found that the annualised share of credit card balances transitioning into serious delinquency, defined as 90 days or more past due, was 6.97%, compared with 6.93% in the same quarter of 2025. For auto loans, the equivalent figure rose from 2.93% to 3.00%.

These figures need to be interpreted carefully. They measure transitions into serious delinquency using outstanding balances, and they should not be confused with the percentage of borrowers who defaulted. They do, however, demonstrate that repayment difficulties remain relevant even when broader economic indicators appear relatively stable.

The Federal Reserve's November 2025 Financial Stability Report also highlighted that consumer delinquency rates remained high by historical standards. It reported that credit card delinquency rates had reached their highest level since 2010 during 2024 before stabilising in the first half of 2025. Much of the earlier increase was associated with borrowers holding nonprime credit scores.

A separate Federal Reserve research note published in November 2025 found that auto loan delinquencies had increased among lower-income households during the third quarter of 2025, even as credit card performance remained relatively stable across several borrower categories.

For banks, the implication is that portfolio averages can conceal important differences between customers. A relatively stable overall delinquency rate may coexist with worsening financial conditions among particular borrower groups. AI could help risk teams examine these differences more frequently and identify where deterioration is concentrated.

A bank that identifies a borrower experiencing financial stress before the first missed payment may have more ways to help. Depending on the customer's circumstances and the lender's policies, those options could include an early conversation about repayment difficulties, financial counselling or an appropriate modification to the repayment arrangement. The benefit of an early warning system depends on what the bank actually does with the information.

AI Could Help Banks Recognise Financial Stress Before a Payment Is Missed

Most people do not move from financial stability to serious delinquency overnight. Financial pressure can develop gradually as expenses increase, income becomes less predictable or existing debt consumes a larger share of available cash.

AI can help analyse combinations of relevant signals that traditional rule-based systems may struggle to evaluate together. For example, a customer might begin carrying a higher credit card balance while their deposit activity becomes irregular. Another customer may experience a sharp reduction in business revenue but continue meeting repayment obligations using accumulated reserves. Both situations could deserve attention, although they would require very different interpretations.

Machine learning models can be trained to identify patterns associated with future repayment difficulties using historical information. Banks can then evaluate whether these models provide meaningful predictive value beyond existing credit risk methods. The important question is how reliably the system identifies genuine deterioration without overwhelming risk teams with false alarms.

This is where model design becomes especially important. An early warning system that flags too many customers will create unnecessary work and may cause banks to intervene where no intervention is needed. A system that is too conservative may miss the very cases it was designed to identify.

A useful approach is to make predictions actionable. The system should help the risk team understand why an account was flagged, which factors contributed to the assessment and whether the situation requires immediate attention or continued monitoring.

A principle we have taken from Phaneesh Murthy's mentorship is that technology needs to fit into the way an organisation actually makes decisions. A predictive score sitting inside a dashboard has limited value if the bank has no process for investigating it or deciding what happens next.

Commercial Lending Could Benefit From Earlier Warning Signals

The opportunity extends well beyond consumer banking. Commercial lenders deal with businesses whose financial positions can change because of delayed customer payments, rising operating costs or weakening demand. A company may remain current on its loan repayments even as the cash available to support those repayments gradually declines.

Commercial real estate offers a particularly useful example. The FDIC's 2025 Risk Review identified commercial real estate among the significant areas of credit risk facing US banks. The review discussed the impact of elevated vacancy rates, slower rent growth and refinancing difficulties on property-level performance. These conditions can affect a borrower's repayment capacity long before a loan enters delinquency.

AI could help banks monitor relevant changes in business performance and collateral conditions across their commercial portfolios. A lender might examine the relationship between rental income, debt service obligations and refinancing requirements. If several indicators begin deteriorating together, the system could prompt an earlier review of the exposure.

This becomes especially useful for banks managing large portfolios with thousands of borrowers. Risk managers cannot manually investigate every change in every business account, so they need reliable ways to prioritise their attention. AI can help identify where a closer examination may be worthwhile, while experienced credit professionals determine whether the concern is justified.

The FDIC's 2026 Risk Review provides some perspective here. It describes generally contained credit risks during 2025, alongside continued weakness in certain commercial real estate and consumer loan categories. The banking industry therefore faces a mixture of broadly stable conditions and specific areas of vulnerability.

This is exactly the kind of environment where more granular risk intelligence could prove useful. A bank needs to understand which exposures are changing rather than assuming that conditions across an entire portfolio are moving in the same direction.

Financial Risk Management Could Become More Connected Across the Bank

One of the challenges with introducing AI into banking is that organisations often approach it one department at a time. The credit risk team may develop a predictive model, the collections team may automate customer prioritisation and the relationship management team may implement its own customer intelligence tools. Each project can improve a particular process, but the bank still needs a way to connect the information.

Consider what happens when an early warning system identifies a borrower whose repayment capacity may be weakening. The risk team receives the alert, but the relationship manager may have information that explains the change. Perhaps the customer has recently changed jobs or is dealing with a temporary interruption in business revenue. The collections team may also need to know whether the account is already receiving support.

Without a connected workflow, different departments could approach the same customer independently, using different information and potentially making conflicting decisions.

AI-powered financial risk management works best when the intelligence is connected to the bank's operating processes. That includes defining who receives an alert, what information they can access, when an account should be escalated and which actions require human approval. The bank also needs to record what happened after the intervention so that it can evaluate whether the system is helping.

In our work under the guidance of Phaneesh Murthy, we have come to understand that enterprise transformation often depends on the systems between departments. An individual application may work extremely well, but its impact will remain limited if the organisation cannot act on the information it produces.

For banks, the opportunity is to create a risk management process in which relevant intelligence can move from detection to assessment and, where appropriate, customer assistance without unnecessary delays.

Predicting Risk Creates New Responsibilities for Banks

There is an obvious concern with systems that attempt to predict financial distress. A customer who is identified as potentially vulnerable may never actually experience repayment difficulties. If the bank responds by unnecessarily restricting access to credit or treating the customer as a higher-risk borrower without adequate justification, the model could cause harm.

Banks therefore need to distinguish between an early warning signal and a decision about the customer.

A model may suggest that an account deserves closer attention. Any resulting action should depend on the quality of the evidence, applicable law and the bank's policies. Customers also need appropriate protections against inaccurate data, discriminatory outcomes and excessive use of personal information.

Regulators have already addressed some of these issues. The US Consumer Financial Protection Bureau has explained that lenders using complex algorithms in credit decisions must still provide specific and accurate reasons when taking adverse action. The CFPB's guidance on complex credit algorithms makes clear that technological complexity does not remove those obligations.

Model governance is equally important. In April 2026, the Federal Reserve, FDIC and Office of the Comptroller of the Currency issued revised guidance on model risk management. It emphasises model development, validation, ongoing monitoring and controls tailored to the risks associated with a bank's models. The revised guidance superseded the earlier SR 11-7 framework.

Banks adopting AI for early warning systems will need to consider whether models remain reliable as economic conditions change. A pattern associated with repayment difficulties during one period may behave differently when interest rates, employment conditions or consumer spending patterns shift.

The question of trust therefore extends beyond whether an algorithm can produce an accurate prediction. Banks need to understand how the prediction was generated, whether it remains valid and what safeguards exist around the decisions that follow.

Early Warning Systems Could Help Banks Understand Wider Financial Risks

Individual borrower monitoring is only one part of financial risk management. Banks also need to understand vulnerabilities developing across the wider economy, particularly when credit growth, debt servicing burdens and asset prices begin moving in potentially concerning directions.

The Bank for International Settlements has spent years researching indicators that can help identify the build-up of systemic financial risk. Its credit-to-GDP gap dataset covers 44 economies and measures the difference between the credit-to-GDP ratio and its long-term trend. The indicator is used to help assess whether credit growth may be becoming excessive relative to economic activity.

Research published by the BIS has also examined debt service ratios as early warning indicators. One study found that credit-to-GDP gaps performed particularly well over longer horizons, while debt service ratios were more informative over shorter horizons. These indicators are designed for macroprudential analysis rather than individual borrower assessment, but they demonstrate how financial risk can build before a crisis becomes visible.

AI could help banks incorporate relevant macroeconomic indicators into portfolio analysis and scenario planning. A lender could examine how changes in economic conditions might affect different borrower groups or business sectors. Risk managers could then use those assessments alongside established stress-testing methods to decide whether particular exposures require additional attention.

The important consideration is that financial risks are connected. A change in interest rates may affect household repayment capacity and commercial property valuations. A weakening business sector may influence employment conditions in a particular region. Banks need to understand these relationships when evaluating how risk may spread across their portfolios.

The Real Measure of AI Will Be What Happens After the Warning

There is a temptation to evaluate predictive AI systems primarily through technical metrics. Banks can measure how accurately a model identifies future delinquency, how frequently it produces false alarms and whether its predictions outperform existing methods. These are important measures because a poorly performing model can create substantial operational and financial risk.

But the business value depends on what happens after the model identifies a problem.

A bank could develop an excellent early warning system and still achieve very little if alerts remain unreviewed or interventions arrive too late. It could also generate significant activity without improving outcomes if employees are asked to investigate large numbers of accounts that do not require attention.

Banks therefore need to evaluate the entire process. How much earlier are relevant risks being identified? Are risk teams able to investigate alerts efficiently? Do appropriate interventions improve repayment outcomes? Are customers receiving assistance that matches their circumstances? Does the process reduce avoidable losses without introducing unfair treatment?

These questions are central to the way we have learned to approach enterprise technology through Phaneesh Murthy's mentorship. A system needs to be judged by the decisions and outcomes it improves. Technical sophistication matters, but the organisation still needs the ability to use that sophistication responsibly.

For banks, that means starting with a clearly defined risk management problem, establishing a reliable baseline and testing whether AI improves performance in real operating conditions. A phased deployment may be more useful than attempting to replace an entire risk management framework at once.

Could the Bank of the Future Intervene Before a Financial Crisis Begins?

The idea of a bank that recognises financial distress before it becomes obvious is compelling, but it needs to be approached with care. Customers do not want every unusual transaction to trigger a warning, and banks cannot assume that financial behaviour will always follow predictable patterns. Economic conditions can change rapidly, while individual circumstances often contain information that an algorithm cannot observe.

Still, the potential benefits are substantial.

The Federal Reserve Bank of New York's figures show that household borrowing and delinquency remain significant issues. The FDIC's risk reviews demonstrate that vulnerabilities can develop differently across lending categories. The BIS's work on early warning indicators shows why financial institutions and regulators have long been interested in identifying risk before it develops into a more serious problem.

AI gives banks additional tools to work on that challenge.

The opportunity is to connect relevant financial information, recognise patterns earlier and make better use of the time available before a problem becomes severe. For borrowers, that could mean receiving appropriate assistance while they still have options. For banks, it could mean better portfolio monitoring and more informed decisions about where to allocate risk management resources.

Our learning under Phaneesh Murthy has reinforced the importance of connecting technology to the business processes it is supposed to improve. In banking, that connection is particularly important because the consequences of a poor decision can affect both the institution and the people who depend on it.

The bank of the future may have a much clearer picture of which financial problems are developing across its customer base. What will matter most is whether it uses that understanding to make timely, fair and responsible decisions.

About the Authors

This article was written by professionals & students mentored by Phaneesh Murthy. A veteran in the field of marketing Phaneesh Murthy acts as a guide & mentor to professionals seeking to upskill & grow in their careers. You may reach out to the team via this website if you would like to take up Phaneesh Murthy's mentorship