What Does a Decline Cost?
Felix Sifuna | Founder & CEO, PEMiG 8 November 2026
When a loan goes bad, the lender has something to count. There is an outstanding balance, an expected recovery, a provision and time spent trying to collect.
When an application is declined, there is also a record: a score, perhaps a reason code, and the policy that produced the decision. Those records can explain why the institution said no. They cannot, by themselves, establish whether saying no was economically worthwhile.
A lender can know exactly which rule stopped an application without knowing whether a different, affordable offer would have created value.
That question has particular relevance in Kenya this week.
On 3 November, Treasury Cabinet Secretary John Mbadi used the NCBA Group Economic Forum to urge banks to pass on lower interest rates and increase lending to businesses, farmers and MSMEs.[1] The following day, Stanbic Bank Kenya announced an upgrade to its credit management platform with Axe Finance, using AI and data analytics across the lending lifecycle. The stated aims include faster decisions, stronger risk assessment and tighter oversight.[2]
These developments bring two expectations into the same room: extend credit to productive activity, and exercise better control over lending.
A credit committee needs a way to assess both.
Last week, we examined the questions an institution can ask its AI vendor. The next question concerns the institution itself: what economic result is the decision system being asked to improve?
In September’s The Borrowers Your Model Never Gets to Know, we examined a basic limitation of lending data. The lender observes repayment on the loans it makes. It does not observe repayment on loans it never issued.
That makes the phrase “wrongful decline” more demanding than it first appears.
A rejected application is not automatically a missed opportunity. The applicant may have been unable to afford the loan, unwilling to accept the available terms, or costly to serve. The original decision may have been entirely sound.
Even a borrower who would have repaid in full might have generated too little income to cover the cost of lending.
The commercial question therefore requires more than identifying possible good borrowers. It requires estimating the value of a different lending policy, within the institution’s affordability, fairness and risk constraints.
The amount requested is not that value. A declined KSh100,000 application does not represent KSh100,000 of lost revenue.
The relevant calculation estimates loan income over a stated period, deducts funding, operating and expected credit-loss costs, and makes an explicit allowance for capital. It also recognises that lending to one applicant may displace another use of scarce funds.
Money already spent acquiring an applicant belongs in the review of acquisition performance. The approval decision turns on the future costs and cash flows that decision can change.
These distinctions matter because an impressive figure for “missed lending” can conceal a weak case for making any of it.
There is also a question about incentives.
An institution can lower its default rate by narrowing the group it is prepared to lend to. Whether that improves the business depends on what it gives up, what it saves and what it does with the released capacity.
Research published in The Review of Economic Studies in 2021 makes this tension concrete. Studying a high-cost lender in the United Kingdom, researchers used the quasi-random assignment of applicants to loan examiners to examine decisions at the approval margin. They found bias against immigrant and older applicants when assessing long-run profitability, but not under the short-run outcome used to evaluate examiner performance.[3]
The findings concern one UK lender. Their relevance here is the mechanism: the measure used to judge a decision-maker can reward decisions that fall short of the institution’s longer-term objective.
An automated system inherits the same problem if its definition of success is too narrow. Better prediction of a chosen outcome does not settle whether that outcome is the right basis for the lending policy.
Default remains essential to the assessment. So do income, costs, capital, customer outcomes and the time over which they are measured.
A profitable loan can still be unsuitable for the customer. Affordability and fair treatment belong inside the policy being evaluated.
Consider a hypothetical policy review.
Get PEMiG’s stories in your inbox
Join Medium for free to get updates from this writer.
A lender is assessing an alternative that would make 1,000 additional offers within its affordability and risk constraints. It estimates that 40 per cent of applicants would accept. That produces 400 booked loans.
Suppose expected contribution averages KSh2,500 per booked loan after funding, operating costs, expected losses and the chosen capital allowance. The estimated contribution is KSh1 million, before the incremental cost of implementing the policy and any value displaced elsewhere.
Every part of that sentence carries an assumption.
At 25 per cent take-up, holding the contribution per booked loan constant, the estimate falls to KSh625,000. A different mix of applicants could change that contribution too. Higher collection costs or weaker recoveries could consume the margin.
This illustration is hypothetical. The figures describe a possible result under specified assumptions; they are neither PEMiG performance results nor a measured loss on past declines.
That is still useful information. It tells the institution which assumptions deserve attention and how much room there is for error.
It also changes the conversation. “We rejected 1,000 applications” says little about commercial value. “This alternative may create a positive contribution, but the result depends heavily on take-up and recovery costs” gives a committee something it can examine.
The discipline lies in making the uncertainty visible before treating the estimate as a business case.
At PEMiG, we use our causal modelling capability to examine how repayment outcomes may change when the lending decision changes: the amount offered, the term, or the policy applied.
Through CCSE, we compare alternative decisions under explicit causal assumptions, with explanations and decision records that can be reviewed and replayed. This brings the lending action itself into the analysis.
For an application declined at the requested amount, the relevant comparison may include a smaller facility or a repayment structure better aligned with cash flow. Each alternative carries its own expected outcome, cost and uncertainty.
The economic interpretation then draws on the lender’s pricing, funding, servicing and capital assumptions. Our causal analysis informs that calculation; the institution supplies the commercial context and the constraints governing its decisions.
The strength of the resulting estimate depends on the data and assumptions supporting it. Missing repayment outcomes remain missing. A counterfactual estimate makes a conditional claim about an alternative decision, rather than turning that alternative into an observed fact.
For a credit committee, that distinction is practical. It separates conclusions supported by experience from those that require further evidence before a policy changes.
A credible assessment of declines consequently compares policies over a defined population and period.
It records who became eligible, what was offered, who accepted, what was booked and how those loans performed. It follows contribution alongside losses and gives the loans enough time to reveal outcomes that an early repayment snapshot would miss.
The comparison also needs to account for changes in applicant mix and economic conditions. A new policy introduced during a stronger trading season cannot claim all the improvement that follows.
Retrospective analysis can identify where the case is promising and where evidence is thin. Where appropriate, a bounded, governed prospective comparison can test an alternative among eligible applicants, with affordability checks, exposure limits and clear stopping rules.
The institution then has a basis for reporting what it has actually learned: the observed result, the modelled range and the assumptions that could change the conclusion.
Sometimes the evidence will support more lending. Sometimes it will support a different offer. Sometimes it will confirm that the existing declines are protecting the institution and its customers.
All three can be valuable findings.
The pressure to grow credit does not answer these questions. Neither does a faster underwriting platform. Their value depends on the decisions they enable and the outcomes those decisions produce.
A credit committee already has strong reasons to examine the loans that went wrong. It has an equally serious commercial reason to examine the business its policy excludes, with the same care about evidence, cost and uncertainty.
When your institution says no, what evidence tells it that it has made the better economic decision?
Sources
[1] Kenneth Mwenda, “Mbadi urges banks to lower interest rates and increase lending to businesses,” People Daily, 3 November 2026. https://peopledaily.digital/business/mbadi-urges-banks-to-lower-interest-rates-and-increase-lending-to-businesses
[2] Stanbic Bank Kenya and Axe Finance, “Stanbic Bank Kenya Upgrades Credit Platform to Advance Lending Efficiency and Risk Management,” media release, 4 November 2026. https://www.axefinance.com/resources/news-blogs/stanbic-bank-kenya-upgrades-credit-platform/
[3] Will Dobbie, Andres Liberman, Daniel Paravisini and Vikram Pathania, “Measuring Bias in Consumer Lending,” The Review of Economic Studies, 88(6), 2021, pp. 2799–2832. https://academic.oup.com/restud/article/88/6/2799/6342424
