Can Banks Deploy AI Without Creating More Risk?

$12.5 billion. That’s what cybercrime cost the global banking sector in 2024. More than one in four cyberattacks worldwide targets banks. In Africa, one in three reported a breach last year. Since ChatGPT launched in November 2022, AI-augmented attacks have quadrupled. The same computing power fueling defenses is now available to attackers at a fraction of its cost five years ago. But one number may be the most telling: 240 days. That’s the average time between a breach and its detection: eight months during which an attacker is inside a banking system and no one knows.

On April 2nd, more than 1,000 financial executives from the UEMOA zone gathered for the 2026 launch of the COFEB / HEC Paris certification program - a 13-year partnership between the French business school and the training arm of the Central Bank of West African States. For ninety minutes, Marc Israel - former Microsoft CTO for Africa, Aetheis founder, author of 20 books on technology and AI, bank board member - addressed one question: how do you deploy AI without creating more risks than you solve?

The tools that detect fraud are exactly the tools used to commit it.

On the defense side, AI performs. Machine learning systems flag transaction anomalies in under 200 milliseconds. Fraud detection accuracy has risen from 72% to 95%. Incident response time in security operations centers has dropped from four hours to 12 minutes. According to IBM’s annual study, organizations running defensive AI reduce their average cost per breach by nearly $1.9 million.

On the other side, AI attacks - with the same algorithms, the same models, the same tools. Deepfakes first. Three seconds of voice is enough to clone anyone. A few minutes of footage - intercepted on WhatsApp or scraped from LinkedIn - can generate a convincing live video. In 2024, a Hong Kong bank wired $25 million after an employee joined a video call with what he believed was his CFO. Face, voice, expressions: rebuilt frame by frame.

Phishing has mutated. Artisanal campaigns - once detectable by spelling errors and suspicious senders - have given way to generative AI attacks personalized to each recipient’s psychological profile. Tools like WormGPT and FraudGPT, available on the dark web without ethical guardrails, can send 100 million flawless messages in any language. Success rates have climbed from 3–5% to 35–65%.

Adversarial attacks target the models themselves: injecting bias into credit-scoring systems to approve fraudulent applications, structuring transactions to slip under detection models, optimizing transfer patterns to bypass anti-money-laundering systems.

Average cost of a data-poisoning incident: $8 to $15 million. Many institutions still treat AI and cybersecurity as separate tracks: one first, then the other. That’s exactly the mistake.

Running both transformations simultaneously without unnecessary exposure is possible, if you follow a six-phase framework built on one guiding principle: start small, learn, move forward.

Phase 1- Maturity assessment. Four dimensions: existing infrastructure and security, human capital and AI/ML capabilities, data quality and governance, executive committee buy-in.

Phase 2 - Prioritization. Marc Israel structures it around an impact/effort matrix. Quick wins first: AI email security (already built into most office suites) and fraud detection (already embedded in most core banking systems).

Phase 3 - Groundwork. AI governance, data quality, skills development. Governance shouldn’t precede action, it should accompany it.

Phase 4 -The POC. Five non-negotiables: mandatory sandbox (no production data), KPIs defined before launch, red team integrated from day one, 8–12 week hard limit, budget kill switch set at the outset.

Phases 5 and 6 - Deployment and optimization. Limited rollout first, continuous monitoring, quarterly red teaming. Then feedback loops, model drift detection, continuous improvement. Across all six phases, one constant: the human stays in the loop - on dashboards, alerts, credit decisions.

90 Days to First Results

Three months. That’s how long Marc Israel gives any financial institution to lay the groundwork for a secure AI strategy and deliver its first tangible results.

Month 1 - Pure diagnosis. No deployment, no POC. Take stock: which AI systems already exist, what cyber vulnerabilities are known, who are the three or four key people to interview (CTO, CISO, business unit heads). Map the attack surface. Identify the three highest-value quick wins. Build the project team.

Month 2 - Foundations. AI governance in place, ethics charter, risk management policy validated. First POC launched on the quick win from month one. Cybersecurity teams receive initial training. A basic monitoring dashboard goes live.

Month 3 - Acceleration. First POC results, Go/No-Go decision. If Go, the second POC launches. The AI red team holds its first session. A board and regulator report is delivered. The 12-month roadmap is finalized.

Five Fatal Mistakes

Deploying AI without thinking about cybersecurity. The most common error. Teams roll out enterprise AI tools without asking a single question about usage, data flows, or what happens if the model is compromised.

Launching POCs without a deadline. Twelve weeks maximum. Beyond that, no one knows what’s being tested anymore.
Blind vendor lock-in. The question to ask before signing anything: if I want to migrate to a different provider or a local vendor tomorrow, can I?

Ignoring change management. 80% of AI projects fail not for technical reasons, but because human adoption wasn’t managed. Microsoft has documented this for years. Marc Israel sees it firsthand: Copilot goes live, everyone’s enthusiastic for the first few weeks - then usage rates collapse.

Treating compliance as a final step. Integrating regulatory requirements after deployment costs five times more. Whether the regulator is the BCEAO, GDPR, or the EU AI Act, the logic is the same: compliance is built into the architecture from day one, not retrofitted at the end.

What This Changes in Practice

“Everything I’ve told you is true for Africa, and it’s true for the rest of the world.” Bank size no longer matters. A small bank in rural Mali faces the same exposure as a major bank in New York. But the reverse holds too: the same tools, the same framework, the same roadmap are available to all of them.

Marc Israel points to a company he’s been working with for a few weeks. Target: three months. At the first checkpoint, 15 days in, they had already completed a month’s worth of work.

What this makes visible is less a question of resources than of method. The dual transformation - AI and cybersecurity - isn’t reserved for banks with the budget to do everything at once. It’s available to any institution willing to start small, set rules before having all the tools, and treat the human element not as a variable to be optimized, but as the one thing AI will never replace in a decision with real consequences.