The loan origination decision is one of the highest-consequence decisions a retail bank makes. Approve a creditworthy applicant and you begin a relationship with measurable lifetime value. Decline them and you lose the relationship to whoever approves them next. Approve an applicant who should have been declined and you absorb the loss while managing the relationship through delinquency and collections. The credit model is the mechanism that navigates between those outcomes, and the quality of that navigation has a direct and quantifiable impact on portfolio performance.
The economics in the guide are precise. Twenty to fifty basis points in credit loss per point of model underperformance. Better underwriting on a $1 billion portfolio worth $2 to $5 million in avoided losses annually. Those numbers describe the performance gap on the population the model can already see. The larger and strategically more important gap is the population the model cannot see at all.
Where the origination decision breaks down
Most retail bank credit models are built primarily on bureau data: credit scores, payment history, existing obligations, derogatory records. Bureau data is reliable, standardised, and well-understood. It is also a lagging indicator of creditworthiness, and it is absent or thin for a significant portion of the adult population that has a demonstrably stable financial life but a limited formal credit history.
Thin-file applicants include younger adults building credit for the first time, recent immigrants who have not yet established a domestic credit profile, self-employed individuals whose income does not present in standard employment verification formats, and adults who manage their finances primarily in cash or through transaction accounts rather than credit products. These applicants are not uncreditworthy. They are invisible to models built exclusively on bureau signals. The decision the model returns on them is not a risk assessment. It is the absence of sufficient data to make one, which defaults to decline.
The commercial consequence of that default is not neutral. Every creditworthy applicant a bank declines because its model cannot see them is a customer acquired by a competitor with better data. That customer does not return. And the competitor’s model improves with every loan it originates that the incumbent’s model could not assess, because each originated loan is a labelled training example. The data advantage compounds over time in a way that is invisible in any individual credit decision but significant at the portfolio level over a lending cycle.
The two economic leakages
The first leakage is on the bureau-visible population. Even where bureau data is available, rule-based underwriting and models trained on limited feature sets make systematic errors in both directions. They decline applicants who would have performed well and approve applicants whose observable signals did not fully capture the risk. Twenty to fifty basis points in credit loss per point of model underperformance is the cost of the second error type. The approval yield impact of the first type is harder to quantify directly but is captured indirectly in the conversion rate gap between institutions with stronger origination models and those without.
The second leakage is on the thin-file population. A bank that cannot assess thin-file applicants effectively is not making a conservative underwriting decision. It is ceding a segment of the market to competitors with different data strategies. The size of this leakage varies significantly by geography and customer base, but in markets with high levels of financial informality or a young and credit-thin customer base, the thin-file population represents material addressable credit demand that bureau-only models systematically exclude.
The two leakages interact. An institution that improves its model on the bureau-visible population but cannot extend the same improvement to thin-file applicants will see portfolio performance improve on the existing customer base while continuing to cede origination share at the margins. An institution that addresses both simultaneously builds both a better-performing portfolio and a broader and more diversified one.
What AI on alternative data looks like in practice
The practical answer to the thin-file problem is not to replace bureau data but to supplement it with signals that provide creditworthiness evidence where bureau data is absent or insufficient. The most accessible and most predictive of these signals for retail banking customers is transactional behavioural data from the bank’s own accounts.
A customer who has held a current account with the bank for two years, receives regular income deposits at consistent intervals, manages their balance without systematic overdraft, and pays recurring obligations on time is demonstrating creditworthiness through behaviour even if their bureau profile is thin. A model that can read those signals, construct a forward-looking income and cash flow assessment, and produce a credit risk score alongside the bureau score is making a materially better-informed decision than one working from bureau data alone.
The features that support this kind of model are already present in the bank’s own transaction data: income regularity, income stability, expense patterns, balance trajectory, payment behaviour on existing products, and early indicators of financial stress. None of these require external data acquisition. They require a model architecture that can consume the bank’s own data at the point of origination and produce a scored output that complements or extends the bureau assessment for the applicant segment where bureau data is insufficient.
The speed dimension
Decision speed is a separate but equally significant dimension of origination performance. Digital-first lenders operating AI-native credit models can return origination decisions in seconds. Traditional bank origination processes that depend on manual review stages, batch processing, or document verification workflows that are not fully automated return decisions in hours or days.
The competitive impact of that difference is concentrated in the near-prime and thin-file segments, precisely where the model quality argument above is most relevant. An applicant in those segments typically has multiple options available to them, and the institution that provides the fastest credible decision has a structural advantage in conversion. The applicant who begins an origination process with an incumbent bank and receives a next-day decision has often already accepted an offer from a digital-first lender by the time the response arrives.
Speed and quality are not independent. A faster decision from a poorly calibrated model that declines too many creditworthy applicants is not a competitive advantage. But a faster decision from a well-calibrated model that accurately assesses the full creditworthy population is. The two improvements compound together in a way that neither delivers as effectively on its own.
The technology dimension
Loan origination systems at most large retail banks are deeply integrated with the core banking platform. Customer history, account data, existing product relationships, and income evidence are all held in the core banking system, which typically runs on IBM Z. An AI model deployed on the same platform as the origination workflow has direct access to the transactional behavioural data described above without data extraction, movement, or transformation overhead. IBM Machine Learning for z/OS supports Python-based model development with deployment to z/OS, enabling data science teams to build models on extracted development datasets and deploy them into the production origination flow with full access to the on-platform data estate. The speed and governance benefits of on-platform inference are the same as those described in the card authorization context: co-located inference eliminates the latency and coverage constraints that off-platform architectures introduce, while keeping the decision and the data it is based on within the same secure, auditable environment.
What success looks like
The metrics that move in a well-executed origination AI programme are approval yield, first-payment default rate, and risk-adjusted return on the portfolio. Approval yield measures whether more creditworthy applicants are being approved. First-payment default rate measures whether the additional approvals are performing as expected. Risk-adjusted return measures whether the combination produces better portfolio economics than the previous underwriting approach. The three are connected and should be monitored together. An improvement in approval yield that is accompanied by a deterioration in first-payment default rate is not a success. An improvement in approval yield with stable or improving default performance at the same or better margin is.
The programme should define target ranges for each metric before the model is deployed, establish the current baseline, and assign accountability to a named business owner. The credit model is a means to those outcomes. The outcomes are what the institution is investing in.
