FSB guidance on AI in financial services

Three ways to see it

1. 1

   FSB's main concerns about AI in finance cluster around four areas. Procyclicality and herding: if many banks use similar models, they will fail together in a downturn. Third-party dependency: a few foreign cloud and model providers serve much of the world's banking AI, creating concentration risk. Cybersecurity: AI expands attack surface and adversarial vectors. Opacity: complex models are difficult for supervisors to inspect. Each concern translates into a control banks should adopt before regulators require it.

2. 2

   Way one to adopt FSB thinking in a Pakistani bank: stress test your AI in a recession scenario. If five percent of borrowers default in a normal year and 15 percent in a stress year, do the model's approvals shift sensibly under stress, or does it keep approving at peacetime rates? FSB wants supervisors to ask this question; the bank should ask it first.

3. 3

   Way two: map your third-party AI dependencies. Most Pakistani banks have not done this. Which foreign cloud holds the model? Which API is the chatbot calling? Which fine-tuning provider has the rights to your historical data? If three of your top dependencies fail simultaneously, what happens? FSB wants a documented map; the bank should keep it on the same shelf as its disaster recovery plan.

The why-tree

Try this with Claude

AI-edge prompt to try: 'Acting as an FSB-style supervisor, ask me ten diagnostic questions about my Pakistani bank's AI estate, covering herding, third-party concentration, cybersecurity, and explainability. Then summarise my three biggest exposures.' Use as a mock supervision exercise.