{
  "video": {
    "id": "kIlVNGZZr8o",
    "title": "How Agents Cut Hour-Long Data Checks to 5 Minutes - Amol Kapoor & Clifford Ressel, Nori Agentic",
    "duration": 1174,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "kIlVNGZZr8o",
    "title": "How Agents Cut Hour-Long Data Checks to 5 Minutes - Amol Kapoor & Clifford Ressel, Nori Agentic",
    "one_liner": "A Fortune 200 asset manager cut the triage of each changed cell in its security master from about an hour to about 5 minutes of human review. It did this with scheduled cloud agents that run a locally deployed open-source model, search Box, Bloomberg and email for the source of each change, and send a cited keep/discard recommendation to the cell owner in Slack.",
    "summary": "The speakers argue that the surge in private assets (secondaries, land, energy, data centers) breaks the security master. Public assets share identifiers like FIGIs, RICs, ISINs and CUSIPs. Private assets have none: they are 'captured and defined entirely by the documentation', so the data grows almost one-to-one with the assets. Their answer is an 'agentic SecMaster', a database tended by 'cleaner fish automatons'. These are generic cloud agents, customized through markdown instructions, skills and AGENTS.md files rather than custom infrastructure. Each one assembles a cited case for every changed cell, and a human gives a yes or no. They close with three lessons: AI adoption is a collaboration problem, aim for 95% automation with a human in the loop, and agents sharply lower the cost of accidental complexity.",
    "key_points": [
      "Market context: $240B of secondaries volume in 2025, up 50% from 2024, and a forecast of $32 trillion in alternative assets under management by 2030. Their explanation is AI being the main growth area, high-growth companies choosing a Series F over going public, and firms looking for edge outside public markets. They cite Matt Levine's point that you can't simply index when half the market is private.",
      "A security master is the master database of every security or instrument a firm holds. It is the source of truth for risk management, P&L accounting and regulatory compliance, and it separates data producers from data consumers. Even public equities are messy: GOOG and GOOGL are different share classes, and if you index on ISIN you can't join data that exists at the legal-entity level.",
      "Public assets have shared identifiers from data vendors (Bloomberg FIGIs, LSE RICs) and numbering agencies (ISINs, CUSIPs). Private deals such as lending to a power-plant builder or buying land near one have no shared identifiers, no single regulator and no single data-vendor API. Repayment schedules and collateralization live in one-off agreements and amendments, in PDFs, Box folders, Outlook email, Zoom transcripts, and sometimes 'on a napkin'.",
      "Case study: a Fortune 200 New York asset manager ran a day-over-day change analysis on a table of about 50,000 assets with about 200 columns each. Every changed cell went to a human triage team, at roughly an hour each, adding up to 'hundreds or thousands' of hours a month.",
      "The fix: whenever the investment universe changes, a Nori cloud agent spins up. It runs a locally deployed open-source model so the data stays on-prem, and it can use Box for PDFs, public APIs like Bloomberg, and a small segment of email. It finds the source of the change, writes an explanation with citations, recommends keep or discard, and pings the cell's owner in Slack. Human time drops from about an hour to about 5 minutes.",
      "They define cloud agents as coding agents like Claude Code or ChatGPT that run in a remote environment instead of on your laptop. That gives auditability, transparency, observability, and scheduled or triggered runs with no person involved.",
      "Domain expertise comes 'the same way a new analyst does': written English instructions about where documents live, which sources to trust and how to judge a change. A generic agent (Claude Code, Open Code, Pi) is 'skinned' with skills or AGENTS.md files, and most of the customization comes from the configs and credentials placed in the environment.",
      "Their architecture view: the SecMaster agent, the audit agent and the code review agent should share one cloud-agent infrastructure. Instructions, tools, triggers and context belong to the company. Orchestration, security and integrations should be identical across tasks and, in their view, owned by a vendor."
    ],
    "takeaways": [
      "Have agents gather the evidence for a decision (search documents, collect citations, recommend keep or discard) and leave a human to approve with a yes or no. Most of the old hour was manual document search, which agents are good at and humans are bad at.",
      "Aim for about 95% automation and keep a human in the loop on purpose, for accountability. Many more problems become solvable when one human spends 5 minutes than when you demand 100% end-to-end automation.",
      "Don't build a custom one-off agent for each workflow. Use a generic agent with markdown instructions, skills, AGENTS.md files and scoped credentials, all on one shared cloud-agent platform.",
      "Make agents proactive: trigger them on data events or schedules so they come to people with tasks and recommendations, instead of a person deciding each time whether to hand work to an agent.",
      "Point agents at 'accidental complexity' in the back office, meaning low-complexity, high-effort work such as reconciling the different formats teams and vendors use. That frees people for the problems that are essentially complex. For sensitive financial data, consider locally deployed open-source models so nothing leaves the environment."
    ],
    "topics": [
      "agents",
      "cloud agents",
      "finance",
      "data quality",
      "security master",
      "private assets",
      "human in the loop",
      "enterprise ai"
    ],
    "tools": [
      "Nori Agentic",
      "Claude Code",
      "ChatGPT",
      "Open Code",
      "Pi",
      "Box",
      "Bloomberg",
      "Slack",
      "Outlook",
      "Zoom",
      "Bloomberg FIGI",
      "LSE RIC",
      "ISIN",
      "CUSIP",
      "OpenAI",
      "Anthropic",
      "Agentics NYC"
    ],
    "quotes": [
      {
        "text": "The asset is captured and defined entirely by the documentation.",
        "at": "09:42",
        "url": "https://www.youtube.com/watch?v=kIlVNGZZr8o&t=582s"
      },
      {
        "text": "you can take something that used to take an hour and make it just 5 minutes of human work.",
        "at": "14:33",
        "url": "https://www.youtube.com/watch?v=kIlVNGZZr8o&t=873s"
      },
      {
        "text": "the future of enterprise AI is not use a robot on your laptop to solve a problem. It's having the robots come to you with tasks, advice, analysis, and recommendations.",
        "at": "16:24",
        "url": "https://www.youtube.com/watch?v=kIlVNGZZr8o&t=984s"
      },
      {
        "text": "aim for 95% automation and it's okay to keep a human in the loop.",
        "at": "17:53",
        "url": "https://www.youtube.com/watch?v=kIlVNGZZr8o&t=1073s"
      }
    ],
    "words": 4314
  },
  "summary_url": "/#kIlVNGZZr8o",
  "transcript": {
    "html": "/transcripts/kIlVNGZZr8o.html",
    "txt": "/transcripts/kIlVNGZZr8o.txt",
    "vtt": "/transcripts/kIlVNGZZr8o.vtt"
  }
}