{
  "video": {
    "id": "ZMZlAn6Pyfg",
    "title": "Proving Alpha: How a Hedge Fund Measures AI's ROI - Hari Kumar, Contour Asset Management",
    "duration": 1104,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "ZMZlAn6Pyfg",
    "title": "Proving Alpha: How a Hedge Fund Measures AI's ROI - Hari Kumar, Contour Asset Management",
    "one_liner": "Contour Asset Management measures AI's ROI by running live A/B tests on analysts' research workflows, documenting the same earnings-cycle process with and without AI tools, and tracking a 'cost per useful insight' metric, because P&L takes months to show up at a fundamental shop.",
    "summary": "Hari Kumar, head of AI at Contour (a TMT-focused fundamental long/short hedge fund with about 15-20 people and $6B AUM), argues that AI ROI can't be shown immediately in investing, so it has to be measured through the process. Contour documents analysts' pre-AI workflows over a three-month earnings cycle and compares them with workflows that use AI tools added one at a time. It also tracks insights per day, time spent per name and eventual P&L outcomes. He describes agents for idea generation, reviving dormant ideas and timing, as well as consolidating ops and compliance into one in-house tool. Contour pipes existing data subscriptions (AlphaSense, Yipit) into its own platform, which adapts to each analyst. The time lost fixing broken connectors and model calls counts as part of the ROI equation.",
    "key_points": [
      "Contour runs 'live AB tests': for one name over a three-month cycle (one earnings report to the next), they record the analyst's whole no-LLM workflow, data sources and conclusions, then rerun it with AI tools introduced one by one.",
      "Example comparison for expert calls: an agent runs the expert call and the analyst synthesizes the notes, versus the agent also synthesizing the notes before the analyst draws conclusions.",
      "ROI can't be shown immediately because theses take 3 months to a year to affect stock price, so they measure process improvements, e.g. consumer surveys fielded by an agent instead of a human, or faster model updates.",
      "A metric they came up with: 'cost per useful insight'. They track insights generated per analyst per day, time spent on each name, which tools produced the idea and the eventual P&L outcome.",
      "Agents scan analysts' notes, even ideas worked on for just a day, track them against the market, and flag ideas that have become interesting again to the analyst, PM and trader, so a key price window isn't missed. Timing is described as the most important thing.",
      "Alt data like Yipit credit card panel data used to sit in isolation. With MCP/API access it now feeds into agents alongside expert call transcripts and notes.",
      "Ops and compliance workflows are split across Bloomberg and order management tools. Contour is merging order reconciliation, vendor tracking and similar tasks into one internal tool. Agents also track entry/exit times and positions for traders.",
      "The ROI equation also counts the time lost when model calls or connectors break 'every couple days' and need fixing. He says that cost is falling but is still a metric they track."
    ],
    "takeaways": [
      "Before rolling out AI tools, write down a baseline of the existing no-AI workflow (data sources, conclusions, time taken). Then add tools one at a time and compare against it.",
      "When the outcome metric (P&L) lags by months, use process metrics as proxies, such as cost per useful insight, time spent per name and insights per day, and tie them back to outcomes later.",
      "Get more out of existing data subscriptions (AlphaSense, Yipit) by pulling them through MCPs/APIs into one internal agent layer, combined with your own notes and transcripts, and adapted to each user.",
      "Use agents to revive dormant ideas and flag them when market conditions make them timely, not just to generate new ideas.",
      "Include maintenance time (broken connectors, failed model calls, user time spent learning tools) in the ROI calculation."
    ],
    "topics": [
      "roi measurement",
      "finance",
      "agents",
      "ab testing",
      "alternative data",
      "workflow automation",
      "mcp",
      "productivity metrics"
    ],
    "tools": [
      "Contour Asset Management",
      "Yipit",
      "MCP",
      "Bloomberg",
      "AlphaSense",
      "Rogo",
      "Kensho"
    ],
    "quotes": [
      {
        "text": "the challenge with something like this is you can't show ROI immediately.",
        "at": "05:43",
        "url": "https://www.youtube.com/watch?v=ZMZlAn6Pyfg&t=343s"
      },
      {
        "text": "tracking things like a cost per useful insight, which is the metric that we've come up with",
        "at": "11:22",
        "url": "https://www.youtube.com/watch?v=ZMZlAn6Pyfg&t=682s"
      },
      {
        "text": "the timing is actually the most important thing as to when you want to strike",
        "at": "08:36",
        "url": "https://www.youtube.com/watch?v=ZMZlAn6Pyfg&t=516s"
      },
      {
        "text": "where model calls break or the connector breaks down every couple days, and then it's your job to fix it, right?",
        "at": "16:58",
        "url": "https://www.youtube.com/watch?v=ZMZlAn6Pyfg&t=1018s"
      }
    ],
    "words": 3622
  },
  "summary_url": "/#ZMZlAn6Pyfg",
  "transcript": {
    "html": "/transcripts/ZMZlAn6Pyfg.html",
    "txt": "/transcripts/ZMZlAn6Pyfg.txt",
    "vtt": "/transcripts/ZMZlAn6Pyfg.vtt"
  }
}