Nigeria's CBN regulatory sandbox correlates with a 42% fall in fraud losses, an increase in detection accuracy to 92.3% and a 77.7% reduction in detection time. The gains appear linked to sandbox participation and operational agility, but lack of random assignment means causality remains uncertain.

Main Finding

The CBN Regulatory Sandbox deployment of AI-driven fraud detection in Nigerian commercial banks yielded large, measurable improvements: a 41.8% average reduction in fraud losses, detection accuracy rising from 74.6% to 92.3%, and average detection time falling by 77.7%. Participation in the Sandbox was the strongest predictor of these performance gains. Compared with US Federal Reserve pilots, the CBN sandbox delivers an “agility advantage” for innovation but highlights urgent needs for stronger model governance and explainability.

Key Points

• Quantitative outcomes
  • Mean fraud loss reduction = 41.8% (SD = 11.7).
  • Detection accuracy increased from 74.6% pre‑AI to 92.3% post‑AI.
  • Average detection time reduced by 77.7%.
• Sample and stakeholders
  • Survey N = 128 practitioners (71.1% response rate) from 14 commercial banks and 22 fintechs.
  • Roles: 52.3% risk/compliance managers, 29.7% IT/AI specialists, 18.0% senior executives.
  • Sandbox status: 47.7% completed testing, 35.2% ongoing pilots, 17.2% observers/eligible.
• Methods and robustness
  • Convergent parallel mixed‑methods design (quantitative + qualitative integration).
  • Paired t‑tests (pre/post), multiple regression, SEM; assumptions checked (normality, VIF < 5).
  • Reliability: Cronbach’s α > 0.82 across constructs; KMO = 0.874, Bartlett p < 0.001.
• Qualitative synthesis
  • 18 in‑depth interviews (banks, fintechs, regulators) produced five major themes (organizational, technical, regulatory drivers and constraints).
  • Key qualitative insights: sandbox participation accelerates piloting and learning; common barriers are cost, skills, explainability, and regulatory uncertainty.
• Comparative insight vs. US Fed pilots
  • CBN sandbox: faster iteration, local adaptation and inclusion emphasis.
  • US Fed pilots: stronger focus on model governance, explainability, third‑party risk and supervisory integration.
• Limitations
  • Self‑reporting bias, short/early adopter window, limited long‑term post‑exit outcomes.

Data & Methods

• Design: Convergent parallel mixed‑methods; pragmatic/post‑positivist orientation with explanatory sequential element.
• Quantitative
  • Instrument: 47‑item structured questionnaire (5‑point Likert), pilot tested; key domains: AI system characteristics, fraud outcomes, adoption factors, regulatory experience.
  • N = 128; analyses in SPSS 28 and SmartPLS 4; paired sample t‑tests, multiple linear regression, SEM; model fit assessed (R2, adj R2, F).
  • Reliability: Cronbach’s α by construct — AI System Effectiveness 0.91; Fraud Reduction Outcomes 0.89; Adoption Barriers 0.86; Regulatory Sandbox Experience 0.84; Comparative Global Insights 0.87.
• Qualitative
  • 18 semi‑structured interviews (35–55 min), audio recorded, transcribed, coded in NVivo 14.
  • Thematic analysis per Braun & Clarke (2006); produced 14 first‑order codes consolidated into five themes and integrated with quantitative results.
• Ethics and validity
  • IRB approval; informed consent; anonymized, encrypted storage; trustworthiness via member‑checking, thick description, audit trail, reflexivity.
• Timeframe & context
  • Data cover initiatives between Jan 2024 and Nov 2025. Paper submitted Aug 20, 2025; accepted Nov 23, 2025; published Dec 23, 2025.

Implications for AI Economics

• Productivity and cost effects
  • Large fraud loss reductions and faster detection imply direct cost savings for banks (improved net margins) and lower expected losses for customers and insurers. These are quantifiable benefits that can alter banks’ cost–benefit calculus for AI investment.
• Adoption dynamics and diffusion
  • Sandbox participation materially increases successful adoption rates — regulatory experimentation lowers adoption frictions and uncertainty, accelerating diffusion among incumbents and fintechs in emerging markets.
• Market structure and competition
  • Faster, more accurate fraud detection can shift competitive advantage to banks/fintechs that adopt AI early, raising market concentration risks if smaller firms cannot bear implementation costs. Conversely, sandboxes can level the playing field by enabling smaller innovators to prove models.
• Regulatory policy and externalities
  • The tradeoff is clear: agility vs. governance. Policymakers should foster sandboxes to realize innovation benefits but simultaneously mandate model governance, explainability, and third‑party risk controls to limit systemic and compliance externalities.
• Investment in human capital and infrastructure
  • Skills gaps and explainability needs translate into recurring costs (training, model‑risk teams, auditing). Economically, these are necessary investments to realize and sustain the productivity gains.
• Financial stability and systemic risk
  • Improved fraud controls reduce operational and credit loss volatility, potentially lowering capital cushions needed for fraud risk. However, widespread deployment of similar AI architectures without governance could generate correlated failure modes — a systemic risk channel that supervisors must monitor.
• Scalability and transferability
  • Results suggest high returns to scaling in similar emerging‑market contexts with high digital transaction growth, but contextual adaptation (local data, languages, fraud typologies) is essential; blind transfer from advanced‑economy pilots (e.g., US Fed) is incomplete without governance adjustments.
• Research and measurement needs
  • Future economic work should quantify long‑run ROI, distributional impacts (consumer fees, access), and second‑order effects (insurance premia, AML compliance costs), and model systemic concentration/externality risks from large‑scale AI adoption.

Bottom line: This study provides strong empirical evidence that sandboxed AI fraud detection can deliver substantial cost‑saving and efficiency gains in emerging‑market banking, but reaping these economic benefits at scale requires parallel investment in model governance, explainability, skills, and regulatory oversight to manage new systemic and distributional risks.