Supervisors expect the fallback to work.

What regulators are asking of AI in critical services, mapped to what Agent Down showed when that fallback was tested.

REGULATORTHE REGULATORY SIGNALWHAT AGENT DOWN REVEALED

Financial Stability BoardGlobal · 15 Oct 2026Sound Practices for Responsible AI Adoption

The FSB identifies third-party dependencies, business continuity, human oversight, and AI-related technology risks as central considerations for financial institutions adopting AI.

The simulated loss of a critical AI provider exposed how quickly operational dependencies can become recovery challenges, particularly when automation has replaced human processing capacity.Read the FSB report →

MASSingapore · 7 Oct 2026

Contingency plans for “potential failures, unexpected behaviour of third-party AI, or discontinuing of support by vendors”.

The scenario trigger was exactly this: the vendor went into administration. No shared recovery strategy emerged.Read the MAS guidelines →

European Supervisory AuthoritiesEuropean Union · 31 Jul 2026Frontier AI and ICT Risk

The EBA, EIOPA, and ESMA call for stronger governance and a consistent supervisory approach to managing the cybersecurity and operational risks of frontier AI models.

The scenario illustrates how reliance on a common AI provider can create concentrated operational exposure, requiring coordinated response and tested contingency arrangements.Read the joint supervisory statement →

OSFICanada · Jul 2026

“Test AI failure and outage scenarios and establish manual fallbacks and continuity measures for AI-supported business processes.”

The manual fallback depended on people, and 70% of them were gone.Read the OSFI bulletin →

FCAUK · 6 Jul 2026

Agentic AI moves firms from assistance to delegation. Staff shift from decision-makers to supervisors.

The scenario's 70% headcount cut is that shift. When the agent failed, the remaining team could not absorb the volume.Read the Mills Review →

APRAAustralia · 30 Apr 2026

“Where AI supports critical operations, credible fallback processes are required.”

Only 16% were confident of holding minimum service levels.Read the APRA letter →

OCC / Fed / FDICUS · 17 Apr 2026

Interagency model risk guidance.

The failed agent was a third-party AI model running fraud verification and sanctions screening. When it went offline, no single recovery strategy emerged and only 16% were confident of holding minimum service levels.Read the interagency guidance →

Quotes from primary sources. Agent Down responses, n=31. Updated October 2026.

Get new signals as they land.

EmailSubscribe →