A new report by iluminr, The State of Resilience in Agentic AI Adoption, found that only 16% of participants in a simulated AI vendor failure were confident that the bank in the scenario could maintain minimum service levels within its impact tolerances. The report highlights the risk of relying on a workforce for recovery after reducing its capacity.
The report is based on a microsimulation conducted in September 2026. The ‘Agent Down’ exercise involved 31 resilience, risk, and business continuity practitioners from 25 organizations across six countries, most from banking and insurance. Participants responded in a personal capacity. They navigated a scenario in which a bank’s critical agentic AI vendor went into administration. Its platform handled fraud verification and sanctions screening at substantially higher throughput than the previous workforce. A year earlier, the bank had cut headcount in those processes by about 70%.
The report identifies four findings about critical AI dependencies:
- Low confidence in service continuity. Only 16% of participants were confident they could maintain minimum service levels within impact tolerances after losing the AI vendor. None selected ‘very confident’.
- Persistent reliance on people for recovery. Of the 31 participants, four selected rehiring as their preferred recovery strategy. A further eight of the ten who proposed an alternative strategy relied on workforce measures, including recalling retired staff or redeploying employees. Together, these responses accounted for 39% of participants.
- Earlier impact tolerance breaches anticipated. In the scenario, 77% of participants expected the loss of AI processing capacity to bring forward the point at which the bank would breach its impact tolerances. The report recommends reassessing the assumptions behind those tolerances as transaction volumes and recovery capacity change.
- Limited attention to the earlier workforce decision. Only five of 31 participants explicitly linked the disruption to the decision to reduce headcount a year earlier. Five named HR when asked which additional stakeholders to engage, although a senior HR representative was already on the incident call.
The report concludes that decisions taken before deployment constrained the bank’s recovery options once the vendor failed. It cites financial regulators’ expectations for credible fallback arrangements and recommends testing the operating model before committing to AI investment and workforce reductions.
