How a Hedge Fund Measures AI ROI Without Waiting for P&L

The problem. Hari Kumar, Head of AI at Contour Asset Management, a $6B AUM fundamental long-short hedge fund, faces the same challenge many engineering leaders do: proving ROI on AI spend. His mandate was to show impact across research, trading, operations, and compliance. The core difficulty is that investment ideas take months to materialize in P&L, so immediate ROI measurement is near impossible.

The method. Kumar's team runs live AB tests on analyst workflows. For a specific stock over a three-month cycle, they document the complete process without AI tools, then repeat the same process with AI augmentations—like agent-synthesized expert call notes or automated consumer surveys. By comparing time spent, data sources used, and conclusions reached, they isolate the AI contribution.

The metric. They developed a 'cost per useful insight' metric. Analysts track time spent per name and per insight, and the eventual P&L outcome. This ties process improvement to the firm's ultimate success metric, even if causality is noisy. It also highlights where analyst time is going—a proxy for whether AI is freeing up capacity for higher-value work.

The broader view. Kumar extends ROI thinking beyond research. On the trading side, AI agents monitor entry/exit timing and cross-industry position risk, feeding ideas back to analysts. In ops and compliance, the goal is to unify fragmented tools (Bloomberg, order systems) into a single platform, reducing repetitive work and freeing people to contribute to idea generation.

The hidden cost. Kumar notes that maintaining AI tooling—fixing broken model calls, connectors—detracts from productive time. This 'time detracted' is explicitly tracked as part of the ROI equation. It's a pragmatic acknowledgment that even exciting AI workflows carry operational overhead.

The takeaway. This is a rare, honest account from someone inside a real financial institution. The AB testing framework is transferable to any knowledge-work organization: define a baseline, measure time and outputs, and isolate the AI delta. The emphasis on cost per insight is a sensible proxy when P&L impact is lagging.