Underwriters received a score and a recommendation with no reasoning. They could not explain a rejection to a borrower or a relationship manager, so they overrode the model constantly and inconsistently.
STUDY 19 / 30 · ANONYMISED · NDA · NBFC & LENDING · INDIA · 4,200 CR AUM
NBFC & LENDING · INDIA A credit decision the underwriter can explain to the borrower.
RoleLead designer + researcher
Timeline14 weeks
Team2 designers, 1 researcher, 1 PM, 2 frontend, 1 risk SME
VerticalNBFC underwriting console · secured SME lending
Surfaced factor-level attribution - which inputs moved the decision, in which direction, by how much - and generated a borrower-facing explanation from the same structure.
Override rate fell and became consistent. The borrower explanation reduced disputes and became a regulatory asset under fair-lending expectations.
11%OVERRIDE RATEfrom 38%
0.81OVERRIDE CONSISTENCY (INTER-RATER)from 0.34
40 secDECISION EXPLANATION TIMEfrom 12 min
THE ARGUMENT
Why the obvious solution was wrong.
The study matters because the product problem was reframed before the interface was polished.
An underwriter who cannot explain a decision will not defend it. The console gave a score and a recommendation, so every rejection required the underwriter to reconstruct a rationale from raw data under time pressure - and when they could not, they overrode. The override rate was not a measure of model quality. It was a measure of how unexplainable the model had been made by the interface in front of it.
Factor attribution renders which inputs moved the decision, in which direction and by how much, ordered by contribution and grouped into categories a borrower would recognise - repayment history, current obligations, collateral, business vintage. The same structure generates the borrower-facing letter, so what the underwriter sees and what the borrower receives cannot diverge. Overrides fell to 11% and inter-rater agreement rose sharply, because underwriters were now disagreeing with a specific factor rather than with an opaque number.
THE INTERFACE CRAFT
The interaction, rendered as a working product surface.
The specimen below is code-native and uses the study's own design logic. The client interface remains protected.
NBFC & LENDING
DECISION TRACEAPPROVE
5 factors · 2 overrides
DESIGN DECISIONS
Positions we would defend.
Each decision names the principle and the product consequence, not a stylistic preference.
Explainability is an interface problem
The model was explainable. The console was not surfacing it, which produced a 38% override rate misdiagnosed as a modelling failure.
Bind the internal and external view
A borrower explanation generated separately from the underwriter's view will eventually contradict it, which is a regulatory event.
Structure the override
An unstructured override is lost information. A factor-specific one improves the model.
PRODUCT LEADER READOUT
What transfers, and what should remain specific to this product.
A case study is useful when its operating principle travels without turning the original interface into a template.
Read the operating condition
For NBFC underwriting console · secured SME lending, the transferable lesson is not a copied screen. It is the condition the interface had to make legible: Which inputs moved the decision, direction and magnitude, grouped into borrower-recognisable categories rather than model features. Rebuild that visibility for your own roles, risk, terminology, and operating cadence.
Protect the design rule
The model was explainable. The console was not surfacing it, which produced a 38% override rate misdiagnosed as a modelling failure. Keep that rule in the acceptance criteria, component states, and production QA record so later visual cleanup cannot erase why the interaction exists.
Measure behaviour after ship
The evidence record is 11% for override rate, from 38%. Recreate the baseline and outcome window before rollout, segment the result by role and context, and state clearly what the measure cannot prove.
RESEARCH RECORD
The work behind the interface.
These artefacts connect the final interaction back to the evidence and product model that produced it.
Override forensics
Reviewed 400 historical overrides with the risk team; 71% lacked any recorded reason and could not be reconstructed.
Factor grouping research
Tested model-feature groupings with 14 underwriters and 9 borrowers until both groups recognised the same categories.
Fair-lending review
Ran the borrower explanation output past external counsel against RBI fair-practices expectations.
Inter-rater study
Measured underwriter agreement on 60 identical files before and after; kappa moved from 0.34 to 0.81.
“They can tell the borrower why. That one change ended most of our disputes.”
Chief Risk Officer, NBFC · under NDA
NEXT STUDY · 20 / 30 · NBFC & LENDING