Proving Alpha: How a Hedge Fund Measures AI's ROI - Hari Kumar, Contour Asset Management
How Contour Asset Management, a $6B fundamental long/short technology, media and telecom fund, measures the return on AI spend through live A/B tests of analyst research, a cost per useful insight metric, timing agents that resurface old ideas, and AI workflows across trading, operations and compliance.
Measuring the return on AI spend at a fundamental fund
Contour Asset Management is a fundamental long/short hedge fund with roughly $6 billion in assets under management, about 15 to 20 people, and a focus on technology, media and telecom. Its investors tend to be technology-forward, but the firm’s processes have run largely the same way for years. The mandate for the head of AI has been to show measurable impact across the firm, starting with research and extending into trading, operations, compliance and finance. The hard part is answering a simple question: given the money and time going into AI, what is actually being gained? That applies to P&L, which takes months to show up in a fund that holds positions over long horizons, and to the more immediate questions of whether the team is generating better ideas and working more effectively.
Live A/B tests of analyst research
The approach that has worked best so far is running live A/B tests on research workflows. The firm has about ten analysts, and each covers a slice of the universe, with a single analyst often following 30 to 40 names. For a given name, the team records and documents each analyst’s full process across a three-month cycle, from one earnings report to the next. One arm reconstructs how the analyst would have worked in a pre-AI world without any LLM tools: which data sources they used, how they built the thesis, and what conclusions they reached. The other arm adds the internal tools being built one by one, such as agent-run expert calls and consumer surveys fielded by an agent rather than a person. The key comparison is whether the analyst would have reached the same conclusion either way, and how long manual sourcing would have taken compared with an agentic approach.
Measuring process rather than just P&L
Because fundamental ideas can take six months or a year to affect the stock price, ROI cannot be shown immediately. The measurement is therefore process-based. The questions are whether the consumer survey process is improving, whether model updates are faster, and whether the AI is actually changing how someone works or simply adding throughput to an existing workflow. Each tool is evaluated individually against that standard.
Reducing anchoring with idea-generation agents
The main goal of fundamental investing is to form a thesis, but the lesson so far is not to become anchored to the initial one. Idea-generation agents take the ideas each analyst develops per name and layer on alternative datasets, expert conversations and further conclusions. They push the analyst to test whether an idea will materialize over the next few months or needs a longer horizon, and they help remove the anchors that come from the first hypothesis.
Timing agents and resurfacing dormant ideas
A recurring problem is that an analyst may spend a few days on an idea, then the PM loses interest or other work takes priority, and the idea sits untouched for months, sometimes past the key price window. Agents that track every note an analyst has written, even from a single day of work, can monitor those ideas against market movement and flag when one becomes interesting again. The alert goes not only to the analyst but to the PM and the trader. The longer-term aim is to extend this from idea generation to timing, meaning when to enter and exit a position, including on ideas that aren’t top of mind.
Combining alternative data
Alternative data such as satellite mapping and granular credit card panels is valuable, but each source is limited when it lives in isolation. Feeding a credit card panel such as Yipit into agents through MCP and API access allows the system to relate expert call notes and transcripts to spend data and see how the two interact. Sources that were previously siloed can now be analyzed together.
Time spent and cost per useful insight
Spend on research, model updates and AI tooling is only part of the cost. Analyst time is the other part. The team tracks insights generated per analyst per day, time spent on each name, and eventual P&L outcomes. This produces a metric they call cost per useful insight. Tracing each day’s ideas through to where they landed, identifying which were interesting, and noting which tools produced them shows where analyst time actually goes. Investing has a very clear success metric in P&L, but it takes far longer to materialize than in most industries.
Extending AI to trading, operations and compliance
Better trading, operations and finance free up time for better research, so the firm is applying the same thinking across the business. On the trading side, the team measures entry and exit timing and how positions are spread across industries and names. Agents monitor those names so traders can surface insights and feed ideas to analysts and associates, rather than spending their time only executing orders. On operations and compliance, many workflows sit inside Bloomberg or an order management system, and vendor tools tend to be piecemeal, covering tasks like order reconciliation or vendor tracking separately. The goal is one consolidated platform that minimizes repetitive work across the order flow.
Getting more from existing data subscriptions
Many hedge funds spend on point solutions such as AlphaSense or Yipit, or on holistic platforms such as Rogo or Kensho. Typically, an analyst uses an AlphaSense subscription in its own portal and then carries the data into their models. Platforms like Rogo bring the data into the model itself and allow conversational access to agents. The direction here is to keep using these providers while bringing their data into an internal platform that learns each analyst’s habits and model patterns. Each analyst gets a tailored setup, and the system learns continuously from data and model updates so the value of the existing subscriptions keeps growing.
The hidden cost of keeping tools running
The final piece of the ROI equation is maintenance. Model calls break, connectors fail every couple of days, and someone has to fix them, often when the tech team cannot resolve the issue directly. The time pulled away from analysis to debug these tools counts against the value of each insight, so the team tracks it alongside the other metrics. That time is decreasing as the tooling matures, but it remains part of the calculation. As AI adoption grows, especially at smaller fundamental shops, a single unified layer serving analysts across many tasks is expected to be very valuable.