OpenAI and Two Sigma Discuss Balancing Agent Autonomy with Safety in Financial Applications

At the AI Engineer New York conference, Joseph Kim of OpenAI stated that model capability will soon allow financial analysts to delegate tasks to agents, predicting by 2027 that the financial industry will rely less on Excel and PowerPoint, and agents will be capable of "investment idea generation and trading." Kim emphasized that safe autonomy depends on placing controls and human judgment rather than imposing restrictions, ensuring traceability for financial data. John Feminella of Two Sigma presented five implemented controls, arguing that safety and innovation are not opposites and criticizing "approval fatigue" as a barrier to innovation and a source of risk.