Every customer creates a series of decisions, and the economic outcome of the book is the sum of them.
Who should we acquire? Who should we approve? At what price, and with what limit? Who should be given more? Who is worth retaining, and at what cost? Who should receive which collection treatment?
Each of those has a probability distribution, an economic outcome, an opportunity cost, and a set of alternatives that were available and not taken.
Most lenders optimise these decisions separately. Very few optimise the economics of the sequence.
That was survivable while growth covered losses and the addressable population was largely thick-file. It is not survivable now.
Digital banks, finance companies and BNPL providers are competing for increasingly similar customers, including segments where the bureau provides less complete information.
None of the techniques required to resolve this are new. NPV decisioning, risk-based pricing, cash-flow underwriting, test-and-learn, continuous portfolio action and propensity-driven collections have been standard disciplines in mature consumer credit markets for years.
The problem is that too little of it is running in production in this market.
That gap is the work. Not a transformation programme. Not a strategy deck. The specific, measurable decisions where a book is:
These are not values. They are commitments a client can check us against while the work is running — and they line up with what regulators in both markets increasingly expect from credit models and decisioning.
Every approval, limit, price and treatment has an expected value that can be estimated before it is made. A policy rule is a decision taken once, in the past, for a population that has since changed. Wherever the data allows it, we replace rules with valuations.
Saudi ArabiaSAMA Responsible Lending Principles require creditors to adopt "a clear, transparent and documented scientific method, criteria and procedures to evaluate the creditworthiness" of a customer, and to use "financial models and tools" to measure ability to meet monthly obligations. Model-based decisioning is not an ambition here; it is the stated expectation.
United Arab EmiratesCBUAE Model Management Standards (Notice 5052/2022) apply to all models supporting decision-making, pricing models included. A credit policy rule set is a decisioning model — whether or not anyone has ever called it one.
An intervention not measured against what would have happened anyway is a story, not a result. We design the control before the programme runs, and we report what the control says — including when it says our own work did not pay.
United Arab EmiratesCBUAE MMS requires ongoing performance monitoring against defined metrics and limits, with escalation and reporting to a Model Oversight Committee at least quarterly. It names "cherry-picking, reverse-engineering and positive bias" as grounds for supervisory response. A clean control is the cheapest defence against all three.
Credit models carry the assumptions of the market they were built in. Here the variables that matter can include employer concentration, income continuity, residency, expatriate status, thin-file populations, bureau coverage, transaction behaviour, product structure and Shariah-compliant financing mechanics. These are not adjustments to an imported model. They are part of the model.
United Arab EmiratesCBUAE AI & ML Guidance Note (11 February 2026) requires training data to be "sufficiently accurate, relevant and representative of the customer populations to which the models will be applied." A vendor model trained on another market's population does not clear that bar by default, and the burden of showing it does sits with the institution.
What we build is meant to run, not to be presented. Thresholds, weights and segment definitions are exposed as parameters your own team can move, documented well enough to know what changes when one moves. If changing it requires us, we built it wrong.
United Arab EmiratesCBUAE AI & ML Guidance is explicit that institutions "should not employ AI models that they have no control over," and must retain the immediate ability, with human intervention, to stop using one. MMS adds documentation sufficient for an independent party to assess the model, plus archiving for rollback.
A decision you cannot explain in a sentence to the person it was made about is a decision you do not fully understand yourself. Explainability is not a reporting layer added at the end — it constrains what you build, and it is worth the constraint. In an increasingly automated decision environment, the ability to explain, challenge and override a decision becomes part of the operating model.
United Arab EmiratesCBUAE AI & ML Guidance (February 2026) requires that consumers can request human review or an explanation of an AI-generated decision, with alternative arrangements for customers who do not want an AI decision at all; disclosures in plain language in Arabic and English; periodic bias testing; and human oversight calibrated to risk.
A credit function measured only on losses will produce a book that is smaller and safer than it should be, and will lose the customers worth having to whoever prices them properly. Sometimes the right answer is to decline more. More often it is to approve more. The question is not how to reduce losses. It is what decision maximises expected portfolio value inside the institution's risk appetite and regulatory constraints.
Saudi ArabiaSAMA sets an explicit affordability floor — finance is bearable where total monthly obligations remain below the customer's monthly disposable income, assessed against documented living expenses. That constrains how much can be advanced. It does not settle which customers should be, or at what price.