Why Banks Disable AI Features Before Production

Chief technology officers (CTOs) at banks often describe the same sequence. The bank buys an artificial intelligence (AI) platform and completes a successful pilot. Then the compliance team asks the vendor to disable the generative AI features before production.

The pilot proves that the technology can work. It does not prove that the bank can govern the technology in production.

Most banks use the 2011 Supervisory Letter SR 11-7 for model risk management. The guidance predates large language models (LLMs) that produce variable outputs. Banks must now apply established model controls to systems with different behavior and failure modes.

Some vendors disable AI features by default for banking clients. This choice reduces the vendor’s examination risk. A bank that enables the features must accept the model risk. It must also define validation, monitoring, escalation, and human-review controls.

The Bank Policy Institute has warned that delayed AI guidance can create risk. Banks are not rejecting every AI system because the technology fails. They often pause deployment because they cannot explain the controls to an examiner.

Our advisory work addresses this gap. We have built AI systems at Visa under regulatory review. We now help banks define controls that support production use. A bank does not need to wait for new guidance. It can build a documented framework that works with current requirements.