The Verifiability Gap in Financial AI
Henry Han argues the governance problem for financial AI agents is verification: whether a decision can be reproduced and defended after the fact.
A paper from Henry Han argues that the binding constraint on governing agentic AI in finance is verification. Banks and FinTech firms already hand real decisions to systems that break down goals, call their own tools, and act with little supervision. Those systems are capable enough. The trouble is reconstructing how they reached a decision after the fact, which Han says the field has barely studied.
Han calls the shortfall the "Verifiability Gap": the distance between the verification that delegated authority demands and the explanation a decision actually leaves behind. He tested it across nine model versions, from a three-billion-parameter local model up to a commercial frontier system, and capability did not buy auditability. Provider updates changed how past financial actions came out, frontier models refused temperature and seed controls, and reproduction rates swung from one run to the next. Even orchestration, the way the agents get wired together, turned out to be its own policy layer that shaped the final decision.
Most governance debates get stuck on how capable the systems are; Han shifts the question to what survives the decision, whether you can reproduce it and stand behind it later. A decision no one can reproduce is a decision no one can defend. That reframes governance from a training problem into an architecture problem.