AIRINA's products operate in regulated environments — BCEAO-zone banking, microfinance, mobile-money. Interpretability and fairness are not optional. The regulator asks; the institution must answer; the model has to be defensible on more than predictive metrics.
This 4-day program equips institutional risk, compliance, and policy teams with working knowledge of the regulatory regime AND the technical methods to satisfy it — from BCEAO directives and BCBS principles through interpretability for regulatory review, fairness audits on African data, and the ethical risks of LLMs in institutional contexts. Taught in Cotonou, bilingual, with a visiting policy fellow alongside AIRINA's research team.
Program Overview
Four days, full-day sessions, taught in Cotonou. Mornings on the regulatory text and the methodological case; afternoons on hands-on audit work with real institutional data (under NDA where required). The final deliverable is a model risk management memo defensible to a regulator — not a slide deck, a memo.
Program structure
- Day 1. The regulatory landscape — BCEAO directives on AI in banking, BCBS principles for model risk management, EU AI Act as comparative case. Working session on a real BCEAO directive.
- Day 2. Interpretability & explainability for regulatory review — interpretable-by-design models, post-hoc methods and their failure modes, what holds up under a regulator's questions.
- Day 3. Fairness & disparate-impact audits on African data — methods, what is appropriate to claim, hands-on audit on a real microcredit dataset.
- Day 4. Ethical risks of LLMs and generative AI in institutional contexts — hallucination, privacy leakage, brand and reputational risk. Final memo and defense.
Certificate
Successful completion — attendance plus the final model-risk memo graded against AIRINA's rubric — earns an AIRINA Ethics & Regulation of AI certificate. The memo is the deliverable; the cert depends on it.
Learning Outcomes
By the end of the program, participants will be able to:
- Read a BCEAO or BCBS directive on AI / model risk in the original, identify what it requires of the institution, and assess whether a given model satisfies it.
- Distinguish interpretable-by-design models from post-hoc-explained black boxes — and articulate when each is appropriate under a regulator's questions.
- Conduct a fairness audit on a real institutional dataset — chose the right disparate-impact metric for the question, run it, interpret the result, defend the choice.
- Identify and quantify the regulatory and ethical risks of deploying an LLM or generative-AI system inside a financial institution.
- Write a model risk management memo that an institutional model-validation team or external regulator could act on without further translation.
Program curriculum (4 days)
- BCEAO directives on AI in banking — current text, scope, application
- BCBS principles for model risk management (BCBS 239, related)
- EU AI Act as a comparative reference, where relevant to African institutions
- Working session: dissect a real BCEAO directive against a hypothetical model
- Interpretable-by-design models — GAMs, EBMs, monotonic gradient boosting, decision lists
- Post-hoc methods — SHAP, LIME, counterfactuals — and their well-documented failure modes
- What survives a regulator's "why did the model do this?" — and what doesn't
- Hands-on: build an interpretable credit-scoring model and document its interpretation
- Disparate-impact metrics — demographic parity, equalized odds, calibration — what each one measures
- When to use which metric, and the trade-offs nobody will tell you about in a tutorial
- Hands-on: fairness audit on a real (anonymized) microcredit dataset, with a written audit report
- The politics: who decides what fair means, and how that decision is itself an institutional act
- Hallucination, citation fabrication, privacy leakage in institutional LLM deployments
- Brand and reputational risk — the practical accounting
- Regulatory exposure under emerging AI-act regimes
Final deliverable: a model risk management memo on a real institutional model, defensible to a regulator. Drafted in the afternoon, defended before the cohort at end of day.
Who Should Attend
This program is for institutional staff whose work intersects with AI deployment, model risk, regulation, or policy — particularly in West African finance and adjacent regulated sectors.
- Risk managers, model validation teams, and regulatory affairs at financial institutions in the BCEAO zone and adjacent francophone West Africa.
- Policy analysts in central banks and ministries working on AI / fintech regulation.
- Compliance officers at MFIs, mobile-money operators, and fintechs deploying ML pipelines.
- ML team leads who interface with regulators on behalf of their institution.
- Public-sector analysts working with quantitative methods in regulated domains.
Prerequisites
- Working familiarity with ML concepts at a non-implementation level — what a classifier is, what training data is, what overfitting means. You don't need to code; you do need to read code.
- Institutional or regulatory context — you should have a stake, on behalf of your institution, in being able to answer a regulator on these questions.
Selection
For oversubscribed cohorts, applicants submit a one-paragraph description of a real institutional case from their work (anonymized) that they want to bring to the program. Selection priority is given to applicants from BCEAO-zone institutions and to teams whose case study is most likely to surface lessons useful to the rest of the cohort.
Brochure
The detailed program brochure (PDF, EN/FR) is sent on request — including day-by-day curriculum, the visiting policy fellow's profile, sample audit reports from past sessions, and the cohort calendar.
To receive the current brochure, write to contact@airina.africa with "Ethics & Regulation of AI — brochure request" in the subject.