NovaBridge: SME credit from 11.4 days to 2.8

NovaBridge Capital is a CBN-licensed digital lender in Lagos serving Nigerian SMEs. This is a conceptual case study: a 9-month transformation program, run Agile/Scrum, in which I led the full BA lifecycle across four workstreams redesigning how NovaBridge originated, scored, and disbursed SME loans.

The program covered the entire credit journey: the application funnel, the underwriting engine, the compliance layer, and disbursement over real-time rails.

NovaBridge was losing 67% of SME applicants before they even reached credit assessment. The few who made it through waited an average of 11.4 days for a decision, because underwriting ran entirely on spreadsheets, manual bureau lookups, and email queues.

There was no automated compliance enforcement against CBN guidelines, and a single analyst maintained the credit scorecard in a version-controlled Excel file that no one else could audit. The lender was leaking applicants at the front of the funnel and carrying regulatory and key-person risk at the back.

• CBN affordability and bureau mandatesEvery automated decision had to enforce the CBN affordability gate (DTI below 40%) and the three-bureau pull across CRC, CreditRegistry, and FirstCentral. Compliance had to be encoded, not remembered.
• Governance before automationNo credit decision rule could ship on engineering judgement alone. All 14 IF/THEN/ELSE rules required approval through the Credit and Compliance Design Authority before sprint commitment.
• A live loan book on a legacy coreThe new journey had to create loan accounts in Finacle 10.2 through its Open API layer and disburse over NIBSS real-time rails, without interrupting existing lending operations.
• Human accountability in the loopFull automation of underwriting was not acceptable for the hard cases. The design had to preserve a manual review band with clear override governance rather than a black-box decision.

The program sat between four groups: the credit analysts whose undocumented heuristics were the real scorecard, the Credit and Compliance Design Authority that approved every decision rule, the engineering teams building the LOS rebuild and integrations, and the Steering Committee that received the UAT Completion Report and owned go/no-go.

Externally, the integration landscape spanned Mono for open banking, the three credit bureaus, Smile Identity for biometrics, and NIBSS for disbursement, each with its own contract, sandbox, and failure behaviour to specify.

I ran six weeks of structured discovery, including shadow sessions with credit analysts in Lagos to surface the undocumented heuristics sitting behind the Excel scorecard. Those sessions converted tribal knowledge into rules that could be specified, governed, and automated.

In parallel, I mapped the AS-IS funnel by querying the application database directly in SQL. The data pinpointed the leak: 44% of abandonment happened at the document upload step. That single finding reshaped the journey design; the biggest win was not a faster back office but removing the step where applicants gave up.

Attack both ends of the funnel at once. At the front, replace the 7-step application with 4 adaptive steps behind a pre-qualification gate, and replace document uploads with data pulled through open banking and identity APIs. At the back, replace the spreadsheet with governed automated decisioning: 14 approved credit rules handling the clear cases, a manual review band for the rest.

The manual band itself got leverage: an AI-assisted underwriting layer giving analysts structured LLM output with feature attribution, under an explicit override governance framework, so human judgement stayed accountable while getting faster.

• option 01Automate the scorecard, keep the journeyCheapest scope: move the Excel logic into a decision engine and leave the application flow alone. It would have fixed auditability but done nothing for the 67% of applicants lost before assessment ever started.
• option 02Buy a packaged LOS end to endAn nCino-class platform out of the box. Strong underwriting tooling, but the Nigerian integration surface, Mono, three local bureaus, Smile Identity, NIBSS rails, and the CBN rule set would all be customisation anyway.
• option 03Rebuild the journey and decisioning on a custom LOSThe chosen path: an nCino-equivalent custom rebuild carrying a redesigned 4-step adaptive journey, nine specified API integrations, governed decision rules, and an AI-assisted manual review band.

• Adaptive complexity over uniform simplicityFour adaptive steps are harder to specify and test than seven fixed ones; branching logic multiplies UAT scenarios. The funnel data said the fixed journey was the problem, so the specification burden was the price of the drop-off improvement.
• Governed automation over full automationPushing the manual underwriting rate to zero was technically imaginable and deliberately not the goal. Keeping a 9% review band with AI assistance and override governance preserved accountability for exactly the cases where a wrong automated decision costs most.
• Rule transparency over model opacityEncoding credit policy as 14 explicit IF/THEN/ELSE rules, approved by the Design Authority, traded sophistication for auditability. For a CBN-licensed lender, a decision you cannot explain is a decision you cannot defend.

1. 01. DiscoverySix weeks of structured discovery, including analyst shadow sessions in Lagos to surface the undocumented heuristics behind the Excel scorecard.
2. 02. AS-IS analysisMapped the funnel with direct SQL queries against the application database, isolating the document upload step as the site of 44% of abandonment.
3. 03. TO-BE journey designDesigned the origination journey from pre-qualification gate through to real-time NIBSS disbursement, reducing the application from 7 steps to 4 adaptive steps.
4. 04. Integration specificationSpecified nine API integration points: the Mono open banking connection, the three-bureau pull (CRC, CreditRegistry, FirstCentral), the Smile Identity biometric check, and Finacle loan account creation among them.
5. 05. Decision rule governanceDocumented 14 credit decision rules in IF/THEN/ELSE logic, including the CBN affordability gate (DTI below 40%) and the three-bureau mandate, and took them through the Credit and Compliance Design Authority before sprint commitment.
6. 06. AI-assisted underwriting designDesigned the layer for the manual review band: LLM output schema, feature attribution requirements, and the override governance framework.
7. 07. UATCoordinated a four-week UAT sprint across 87 test cases, managing defect triage in Jira and delivering the UAT Completion Report to the Steering Committee.

The origination journey runs on a custom-rebuilt LOS fronting Finacle 10.2 through its Open API layer. Applicant data arrives through Mono open banking and Smile Identity biometrics rather than uploads; decisioning runs on Experian Power Curve against the governed rule set, with the three bureaus pulled in-flow; approved loans create accounts in Finacle and disburse over NIBSS real-time rails. Nine API integration points were specified end to end, each with validation rules and failure behaviour.

Application drop-off fell from 67% to 39%. Time to disbursement dropped from 11.4 days to 2.8. The manual underwriting rate fell from 41% to 9%, with the remaining referrals supported by the AI-assisted review layer, and SME applicant NPS climbed from +12 to +39. Credit policy moved out of one analyst's spreadsheet into a governed, auditable rule set enforced on every application.

• ›The real credit policy lives in analysts' heads, not the scorecard file. Shadow sessions surfaced heuristics no document captured, and without them the automated rules would have encoded the wrong policy.
• ›Query the funnel before redesigning it. The SQL analysis showed the document upload step was the problem; without that number, the redesign would have optimised the wrong stages.
• ›Get decision rules approved before sprint commitment. Taking all 14 rules through the Design Authority up front meant engineering built against settled policy instead of relitigating credit logic mid-sprint.