Agentic AI can generate costs and decisions faster than any oversight process tracks, and most organizations practicing agentic AI risk management still cannot name who owns the resulting judgment calls or when a system must stop.
The person who designs an agent's spending rules doesn't exist at most firms: they can name a budget holder, but not who decides what it can spend or when it must stop.
A spending cap fixes an invoice, but it leaves the underlying judgment gap in place: whether an agent's confidence was ever calibrated to the outcomes it affects.
The FCA's live testing cohort and the EU AI Act's high-risk obligations converge on a 12- to 24-month window for firms to prove they can govern autonomous systems.
Governance without technical depth is theater. Judging whether an agent's confidence is calibrated to real outcomes requires understanding how it fails, not just how it scores on a benchmark.
METR's 2026 research found maintainers judged AI code contributions on a time horizon roughly seven times shorter than automated benchmarks predicted.
Disagreement between agents trained on different data is signal, not noise: averaging it away discards the clearest evidence of where a decision genuinely needs human judgment.
Agentic AI risk management has a structural gap most organizations haven't filled: the person who designs the rules governing what an autonomous system can spend, which decisions it can make without a human checkpoint, and when it must stop. Eduardo Ariño de la Rubia argues the real risk isn't overspending, it's undelegation, handing decisions to systems before anyone has defined who owns the residual risk.
Spending caps fix the invoice; they leave that judgment gap in place, and no set of guardrails for agentic systems closes it by itself. A 2026 survey found most banking professionals cannot confirm their institution could shut down a malfunctioning model, and emerging model risk management guidance is starting to force the question.
Ariño makes the case that data scientists are the natural occupants of a role organizational theorists named in 1967: the integrator, accountable for outcomes rather than communication. The session closes with six diagnostic questions for calibrated confidence before any agent acts on a consequential decision.
