Company Secretary Governance Risk Scores: Developing an AI-based Corporate Governance Index for South African Companies

Purpose: This study develops and validates an AI-Based Corporate Governance Index (AI-CGI) for South African listed firms, anchored on a Company Secretary Governance Risk Score (CSGRS). It addresses the absence of a publicly reproducible, theory-rooted and machine-learned governance measure tailored to the Companies Act 71 of 2008, the King IV Code and the JSE Listings Requirements, at a time when recurring governance failures continue to undermine investor confidence in the Johannesburg Stock Exchange (JSE).

Design/Methodology/Approach: A mixed-methods design was adopted. Quantitative arm: 132 JSE-listed firms were sampled across the 2019-2024 period (660 firm-year observations). Twenty-two governance variables, drawn from integrated annual reports, SENS announcements and the Companies and Intellectual Property Commission (CIPC) registry, were used to construct a Company Secretary Disclosure Quality (CSDQ) sub-index and a Governance Risk Disclosure (GRD) sub-index. Three AI algorithms (Random Forest, XGBoost, Deep Neural Network) were trained on an 80/20 stratified split with SMOTE oversampling and benchmarked against an Ordered Logistic Regression. Qualitative arm: 24 semi-structured interviews with company secretaries and audit committee chairs were coded by reflexive thematic analysis.

Findings: The Random Forest AI-CGI achieved an out-of-sample AUC of 0.913 and an F1-score of 0.876, outperforming both the XGBoost model (AUC = 0.889) and the Ordered Logistic Regression benchmark (AUC = 0.762). The five strongest SHAP-attributed predictors were company-secretary independence, board diversity, audit-committee financial-expert presence, internal-control deficiencies and related-party transaction completeness. The CSGRS showed strong convergent validity against the Institute of Directors South Africa (IoDSA) Governance Instrument (r = 0.74, p < 0.001).

Practical Implications: The AI-CGI offers regulators, auditors, proxy advisers and boards a defensible and reproducible instrument for the early identification of governance-risk concentration. It enables the CIPC and the JSE to evolve from reactive enforcement toward risk-based supervision, supports audit committees in prioritising assurance resources, and gives institutional investors a transparent, interpretable signal for stewardship engagement under the Code for Responsible Investing in South Africa.

Originality/Value: This manuscript is the first to fuse company-secretary-specific governance attributes with ensemble machine learning for JSE-listed firms. It introduces a domain-anchored Company Secretary Disclosure Quality sub-index that captures governance quality beyond board composition alone and supplies a publicly replicable scoring protocol that converts King IV from principle-based guidance to quantitatively auditable practice.