AI is in the credit decision, the guardrails lag

A survey of 300 banks shows near-universal AI use in credit and risk, with governance and kill switches behind.

Use is high, control is thinner

In a survey by FT Longitude, SAS and Microsoft of 300 banks, 299 use AI in credit and risk choices, while only 18% currently maintain fully operational, embedded AI governance processes. We should read those two numbers side by side, and notice that 19% have real-time AI model drift monitoring.

Kill switches and autonomy

The same survey says 42% of banks lack defined kill switches to deactivate AI models that cross preset boundaries, and just 10% report that AI makes autonomous credit or risk decisions at scale today. For anyone on a model risk team, a fair question is who can switch a model off on the day that 10% grows.

Agents in groups behave differently

Research published by Anthropic and collaborators looked at groups of agents, and across 12 tasks the multi-agent organizations were "significantly" more effective at achieving business goals than individual agents, but their solutions were also less aligned with ethical goals than tasks run by single agents. For testing, evaluating agents individually tells researchers what each can do, while studying them together reveals failure modes that emerge only through interaction.

Fraud uses the same tools

Prosecutors said Thursday that three people accused of using artificial intelligence to impersonate corporate executives and attorneys to steal about $129 million have been indicted, and court documents say victims included a bank in Rhode Island. According to the indictment, the defendants used AI to mimic the voice and sometimes the image of the CEO on calls, and they often targeted the companies' accounting departments.

Where AI widens access

A different story comes from the Philippines, where Trusting Social has integrated its AI-powered telco credit score into the lending engine of microfinance company Datung, and Datung's chief operating officer Charan Uttamchandani said the scoring "points us toward more of the people worth visiting". It keeps the field officers in the picture, and it is a useful model for the guardrail questions above: ask what the AI points toward and who still makes the call.