For decades, financial institutions have operated within well-established risk and compliance frameworks. Anti-Money Laundering controls, Know Your Customer processes, fraud detection rule engines, and information security programmes formed the backbone of how banks managed exposure and met regulatory obligations.
These frameworks were designed for a world where risk was transactional, decisions were human-led, and systems were deterministic. A rule triggered an alert. An analyst reviewed a case. A decision was documented. A regulator examined the record.
AI has broken this model, not because AI is inherently dangerous, but because it operates in fundamentally different ways from the systems these frameworks were built to govern.
Four Ways AI Changes the Risk Equation
Decisions are increasingly model-driven
Credit assessments, fraud flags, customer risk scores, and even AML alerts are being influenced or made directly by AI models. The decision trail is no longer a human choosing between options, it is a model producing an output based on training data, weights, and features that may not be fully understood by the institution operating it.
Data flows dynamically across systems and external tools
Employees use AI tools, sometimes officially sanctioned, sometimes not, to process customer data, draft reports, summarise cases, and generate content. Every time sensitive data enters an external AI system, it moves beyond the institution's direct control. Most banks have no visibility into how frequently this happens, or what data is involved.
Employees interact with external AI platforms daily
The frontier of AI risk is not in the data centre. It is on the desktop. A compliance officer summarising an STR in a general-purpose AI assistant. A fraud analyst pasting transaction records into a chatbot to speed up their notes. A relationship manager uploading a customer profile to generate a meeting summary. These are not edge cases. They are standard operating behaviour across most financial institutions today.
Threat actors use AI to scale attacks
Fraudsters and financial criminals have access to the same AI capabilities as legitimate institutions, and they are using them. Deepfake audio for social engineering. Automated credential stuffing at machine scale. AI-generated synthetic identities that can operate undetected within a bank for months. The attack surface has expanded significantly, and the speed of exploitation has accelerated.
The New Risk Category
What emerges from this analysis is a new category of institutional risk, one that sits at the intersection of data, models, behaviour, and compliance. It does not belong neatly to any single function. It is not a technology problem that IT can solve. It is not a compliance problem that the MLCO can own alone. It is not a security problem that the CISO can contain in isolation.
This is the problem that the BAGS™ Framework was designed to solve. In the next article in this series, we examine exactly where current governance structures are failing, and why the gap is larger than most institutions realise.