{
  "video": {
    "id": "RGiXcVxSD3s",
    "title": "Chat and citations won't save your vertical AI - Atul Ramachandran, Filed Inc",
    "duration": 912,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "Online"
  },
  "analysis": {
    "video_id": "RGiXcVxSD3s",
    "title": "Chat and citations won't save your vertical AI - Atul Ramachandran, Filed Inc",
    "one_liner": "Filed's CTO argues that chat (synchronous) and citations (verification pushed back onto the user) break the very promise vertical AI sells, and that products should instead be built as a conveyor belt the user supervises — delegate, teach, monitor, control — and measured in weekly active sessions rather than weekly active users.",
    "summary": "Atul Ramachandran, CTO/co-founder of Filed (AI for US tax professionals, $17M+ raised), says chat and citations are good but insufficient: chat is synchronous so customers can't leave the platform, and citations hand the verification burden back to the customer, which breaks the 'agents do the work while you sleep' pitch in healthcare, legal and taxes. He frames a third abstraction layer — agentic delegation — after physical branches (bottleneck: employees) and digital self-service (bottleneck: users), where the bottleneck disappears because agents work when the user isn't there. His concrete build spec is a conveyor belt with users as supervisors, requiring four components: find delegatable long-running tasks, skills that teach the agent the user's way of working, monitoring via task lists and traces, and control via pause/intervene. He closes by arguing the metric must change from weekly active users to weekly active sessions.",
    "key_points": [
      "Chat is synchronous: once the request is typed the user is waiting on the agent, so the medium itself prevents the customer from leaving the platform and going to do their own work.",
      "Citations do reduce hallucinations (forcing the agent to present references makes it more accurate) but they push the verification burden back onto the customer, who now reviews the agent's work one item at a time — extra work in exactly the verticals that sell time savings.",
      "Three levels of product abstraction, illustrated with a bank: physical branch (bottleneck = number of employees), digital/mobile self-service (bottleneck = number of users), and agentic delegation (bottleneck removed, because agents work while the user has gone to sleep).",
      "The build model: the product is a conveyor belt, AI agents are the workers, and users are supervisors rather than operators — 'design for delegation, not participation.'",
      "Four required components: (1) find delegatable tasks, (2) let users teach the agents, (3) monitor, (4) control/intervene.",
      "Delegation targets tasks that take users more than a couple of hours and are repeatable; at Filed they identified three tasks in the tax workflow taking over an hour each, and those became long-running background agents.",
      "A background agent gets you 80–90% there; the remaining 20% is the quirks of how a particular user or firm does the work, captured as skills — and skills must be captured automatically from product usage, not via a separate interface where users go and author skills ('that won't work').",
      "Monitoring means a task list plus traces: Filed traces back every single value the agent produced, in a format users can easily read — 'this is where the trust is built' and where most complaints get addressed.",
      "Control means pause the belt, fix the problem, restart it. Filed pauses whenever the agent is about to make an assumption; the user tags the agent Slack-style and replies with how to handle that conflict.",
      "Level 2 features don't disappear — physical branches survived mobile banking — and irreversible or dangerous actions need a plan approved first (before data entry into tax software that could erase existing data, Filed presents a plan for approval).",
      "Metric change: weekly active users doesn't translate to delegation. Use weekly active sessions — a task completed by a human or agent even when the user isn't on the platform. The aim is WAU going down while weekly active sessions go up (not to zero)."
    ],
    "takeaways": [
      "Audit your users' workflow for tasks that take more than an hour or two and are repeatable, and turn those into long-running background agents — that's where the core value is, not in a better chat box.",
      "Build automatic skill capture by watching product usage rather than shipping a separate 'create a skill' interface; the last 20% — the user's own conventions and quirks — is what makes the output actually usable.",
      "Ship a task list and per-value traces so the supervisor can see what every agent did without re-verifying citation by citation; treat this as the trust surface, not a nice-to-have.",
      "Add pause-and-resume control: pause the run whenever the agent would otherwise make an assumption, and let the user reply in-thread; keep your level-2 (self-service) features so users know they can take the wheel.",
      "Present an approval plan before any irreversible or destructive action (e.g. writing into a system that can overwrite existing data).",
      "Replace weekly active users with weekly active sessions as the north-star metric, and expect WAU to fall as trust and delegation rise."
    ],
    "topics": [
      "vertical-ai",
      "agents",
      "background-agents",
      "product-design",
      "delegation",
      "skills",
      "observability",
      "metrics"
    ],
    "tools": [
      "Filed",
      "Slack"
    ],
    "quotes": [
      {
        "text": "This synchronous medium does not allow the customers to leave the platform and go and do their work.",
        "at": "02:31",
        "url": "https://www.youtube.com/watch?v=RGiXcVxSD3s&t=151s"
      },
      {
        "text": "Essentially, design for delegation, not participation.",
        "at": "06:16",
        "url": "https://www.youtube.com/watch?v=RGiXcVxSD3s&t=376s"
      },
      {
        "text": "So ideally it should go like you pause the belt, fix the problem, you start it back up again.",
        "at": "11:21",
        "url": "https://www.youtube.com/watch?v=RGiXcVxSD3s&t=681s"
      },
      {
        "text": "your actual aim should be the weekly active users go down while weekly active sessions go up",
        "at": "13:57",
        "url": "https://www.youtube.com/watch?v=RGiXcVxSD3s&t=837s"
      }
    ],
    "words": 3353
  },
  "summary_url": "/#RGiXcVxSD3s",
  "transcript": {
    "html": "/transcripts/RGiXcVxSD3s.html",
    "txt": "/transcripts/RGiXcVxSD3s.txt",
    "vtt": "/transcripts/RGiXcVxSD3s.vtt"
  }
}