New York Fed risk chief: automating junior work may thin future judgment
Serguey Asael Shinder2 min readMihaela Nistor, chief risk officer of the Federal Reserve Bank of New York, used a speech at the Risk Live North America conference in New York on 24 September to describe a risk from artificial intelligence that no control is built to catch. She stressed that the remarks were her own views.
Her framework has three parts: task exposure, which says what can change; deployment friction, which says how fast; and organizational redesign, which determines what happens next. Exposure is not displacement, she said, and capability is not deployment. Large institutions absorb technology slowly for good reasons, including regulation, liability, legacy systems and data quality, and much of that friction is protective. The problem is that different parts of an institution move at different speeds: a unit can automate junior analytical work before the jobs that depend on it are redesigned, or deploy autonomous agents before governance understands the resulting concentration of decisions.
Her main example was training. Institutions have long produced experts through repetitive, lower-level work: junior analysts building reports, junior underwriters reviewing files, control testers walking process maps line by line. That work taught which numbers looked wrong and when to escalate. AI can absorb much of it, delivering short-term efficiency while weakening the pipeline that produces experienced judgment. The erosion, she said, will not show up in a loss event or trip a control; it will show up five to seven years from now as thinner leadership benches and weaker institutional memory.
What it means
Supervisors usually talk about AI risk in terms of models: accuracy, bias, explainability. This is a different category, a risk to the institution rather than to any single output, and one that existing frameworks do not measure. Coming from the risk chief of the bank that runs the Fed's market operations, it is a signal of what supervisors may start asking about: not only whether automated work is done correctly, but who will still know how to check it.