Key Takeaways
- • Contour Asset Management, a roughly $6 billion fundamental long/short hedge fund, is running live A/B tests to determine whether AI genuinely improves investment research rather than merely accelerating it, according to CEO-level remarks by Hari Kumar at AI Engineer.
- • The test benchmarks a fully manual three-month cycle from one earnings report to the next against the same coverage name with AI workflows added incrementally, measuring sourcing time and whether conclusions change.
- • Contour tracks a metric it calls "cost per useful insight," combining insights per analyst per day, time per name, and eventual P&L, which Kumar noted arrives over months rather than days.
- • The fund is consolidating fragmented data, including Yipit credit card panel data and AlphaSense call transcripts, into an internal platform via MCP and API access, positioning it as an alternative to bundled systems such as Rogo or Kensho.
- • No outcome data has been disclosed yet; Kumar acknowledged that maintenance of model and connector breaks, required every few days, still offsets some efficiency gains.
Summary
Hari Kumar told AI Engineer that Contour Asset Management, which manages about $6 billion, is running live A/B tests to establish whether artificial intelligence improves the quality of research or only speeds it up. The firm has 15 to 20 staff, including roughly 10 analysts who each cover 30 to 40 names, so Kumar said any benefit must appear first in the research process before it shows up in performance.
The methodology is intentionally straightforward. According to Kumar, the team first documented a complete three-month cycle, spanning one earnings report to the next, without any AI tools, logging data sources and conclusions. It then reran the same names with new workflows added one at a time. Examples include agent-run expert calls with and without agent synthesis, and agent-fielded consumer surveys compared with human-fielded ones. The aim is to measure how long sourcing takes and whether conclusions change.
The central yardstick is what Contour calls "cost per useful insight." Kumar said the metric blends insights generated per analyst per day, time spent per name, and eventual P&L. At a fundamental long/short shop, he noted, P&L arrives over months rather than days, which makes the feedback loop slow.
To feed the system, Contour is consolidating fragmented data inside an internal platform. Kumar described ingesting credit card panel data from Yipit and call transcript data from AlphaSense via MCP and API access, so agents can combine these datasets with expert notes instead of analysts working across separate portals. He framed this as an alternative to purchasing a bundled system such as Rogo or Kensho, arguing the internal build learns each analyst's modeling patterns. The same consolidation logic extends to trading, where agents track entry and exit timing and cross-industry exposure, and to operations and compliance, where the goal is to replace piecemeal Bloomberg and order management tools with a single reconciled workflow.
Kumar was candid that outcomes are not yet demonstrated. Idea generation is slow to validate, and timing often matters more than the idea itself, so agents also monitor stale notes and flag shelved theses that become relevant again. He also identified a cost that ROI calculations often omit: the time spent repairing model or connector failures every few days. He said that burden is declining but still offsets gains. Absent disclosed changes in hit rate, time saved per insight, or cost per insight over multiple quarters, the account describes a process improvement rather than proven returns.
The methodology is intentionally straightforward. According to Kumar, the team first documented a complete three-month cycle, spanning one earnings report to the next, without any AI tools, logging data sources and conclusions. It then reran the same names with new workflows added one at a time. Examples include agent-run expert calls with and without agent synthesis, and agent-fielded consumer surveys compared with human-fielded ones. The aim is to measure how long sourcing takes and whether conclusions change.
The central yardstick is what Contour calls "cost per useful insight." Kumar said the metric blends insights generated per analyst per day, time spent per name, and eventual P&L. At a fundamental long/short shop, he noted, P&L arrives over months rather than days, which makes the feedback loop slow.
To feed the system, Contour is consolidating fragmented data inside an internal platform. Kumar described ingesting credit card panel data from Yipit and call transcript data from AlphaSense via MCP and API access, so agents can combine these datasets with expert notes instead of analysts working across separate portals. He framed this as an alternative to purchasing a bundled system such as Rogo or Kensho, arguing the internal build learns each analyst's modeling patterns. The same consolidation logic extends to trading, where agents track entry and exit timing and cross-industry exposure, and to operations and compliance, where the goal is to replace piecemeal Bloomberg and order management tools with a single reconciled workflow.
Kumar was candid that outcomes are not yet demonstrated. Idea generation is slow to validate, and timing often matters more than the idea itself, so agents also monitor stale notes and flag shelved theses that become relevant again. He also identified a cost that ROI calculations often omit: the time spent repairing model or connector failures every few days. He said that burden is declining but still offsets gains. Absent disclosed changes in hit rate, time saved per insight, or cost per insight over multiple quarters, the account describes a process improvement rather than proven returns.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.