Underwriting Model Validation and Responsible AI in Nigeria: Review Guide

An underwriting model should be evaluated for its intended use, population and operating conditions. A scorecard that performs well in one setting may not support the same decisions elsewhere. Wigmore Trading accepts model-requirement, procurement and specialist-coordination enquiries in Nigeria. Development, validation, formal audit and professional opinions require a defined scope with competent providers; this page does not certify a model or approve individual applicants.

Separate development, validation and audit

Development builds or changes a model. Validation assesses whether the model is suitable for its stated purpose. Audit examines a defined subject against specified criteria and can have a different independence requirement. Ask what work is actually included, who performs it and what limitations apply. Do not describe a vendor demonstration or simple test as an independent algorithm audit.

Data quality before algorithm claims

Document source coverage, missing records, labels, timing and changes in the business environment. Check that information available only after the outcome has not distorted evaluation. Ask qualified reviewers to assess development and testing methods for the actual use case. More data is not automatically more relevant, fair or lawfully usable.

Bias and explainability need context

A fairness review should define the harms, affected groups, lawful evidence and appropriate evaluation methods. One overall accuracy figure cannot establish fair treatment. Explanations should be useful to the intended reviewer and reflect the model’s actual limitations. Do not claim a system is unbiased or compliant merely because a dashboard offers a reason code.

Use risk-management references appropriately

The NIST AI Risk Management Framework is a voluntary reference for managing AI risks, not a Nigerian underwriting licence or a product certificate. The NDPC’s published report discusses human review in the context of consequential automated decisions. Qualified advisers should assess applicable obligations for the particular institution and use.

Monitoring after deployment

Agree how model performance, data changes and operational errors will be monitored. Define triggers for investigation, restricted use, recalibration or retirement. Keep model versions, approval records and exceptions traceable. An initially successful validation does not establish that a model remains suitable indefinitely or that automation removes credit or insurance risk.

Compare review proposals

Request the scope, independence, evidence needs, methods, deliverables and follow-up arrangements. Separate a technical model assessment from legal advice, regulatory assurance and financial-statement audit. Confirm how confidential data and intellectual property are protected. A fixed fee is meaningful only when the review boundary and information requirements are clear.

Discuss a model-review requirement

Tell Wigmore the institution type, model purpose, population, current stage and desired deliverables using a redacted summary. We can help organise the support enquiry and specialist dependencies. Do not send borrower records, model credentials or confidential datasets in the first message.