{
  "video": {
    "id": "_cVfz88_j7A",
    "title": "Can Oncology Workflows Run Without Human Touch? - Anant Shankhdhar, Risa Labs",
    "duration": 1001,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "Online"
  },
  "analysis": {
    "video_id": "_cVfz88_j7A",
    "title": "Can Oncology Workflows Run Without Human Touch? - Anant Shankhdhar, Risa Labs",
    "one_liner": "A Risa Labs AI engineer explains how they removed the human reviewer from parts of oncology prior authorization by stacking deterministic checks, multi-source evidence reconciliation, and four agents (eligibility, auth status, medical necessity, submission), escalating only what falls below a confidence threshold.",
    "summary": "Anant Shankhdhar describes Risa Labs' prior-auth pipeline for cancer drugs — order intake, eligibility and benefits verification, deciding whether each drug is NAR (no auth required), auth-on-file, or auth-required, then submission — where bots already ran the steps but a human had to review every order before submission. His task was to identify, confidently, which orders could go straight to submission untouched, so confidence became the key metric. The answer was layered: a unified payer-connection service with a deterministic decision engine, LLM-generated RPA configs plus a self-healing loop to make portal automation scalable, multi-source evidence (authorization letters plus a payer-rule SQL knowledge base) to make LLM-extracted auth statuses trustworthy, and a medical necessity agent that attaches a confidence score and escalates only the cases a clinician actually needs to see.",
    "key_points": [
      "The workflow is four agents: EV agent (eligibility and benefits verification), auth agent (drug status), necessity agent — 'the clinical brain of the system' — and submission agent; each drug falls into NAR (no auth required), auth on file, or auth required.",
      "Insurance data is 'scattered across dozens of portals, APIs and documents', so they built a unified service that connects to different payer sources and emits a normalized coverage result, then a deterministic decision engine that stops orders whose coverage isn't active — no human needed for those.",
      "A coverage orchestrator routes each patient down either an API path or an RPA path, both producing the same fixed coverage-result format.",
      "Custom per-portal integrations don't scale, so they built a repository of custom and popular RPA actions plus LLM-based config generation that assembles a portal's automation config, cutting development time significantly; a self-healing loop detects and mitigates automation breakages during production hours.",
      "The first attempt at auth status was a plain LLM extraction over patient notes (defaulting to 'auth required'), but the notes often lacked the data and LLM extraction is non-deterministic, so outputs 'cannot be blindly tested' — it improved efficiency without removing the human.",
      "The fix was multi-source evidence: prior authorization letters from the patient's earlier runs confirm auth-on-file, and a payer rule knowledge base (a SQL database built from portal checks, LLM extraction over time-scoped payer documents, and historical knowledge of how an organization treats a drug) confirms NAR. Where two sources concur, the status is asserted with confidence; everything is configurable so cases can be plugged manually.",
      "Because some orders contain only NAR and already-authorized drugs, those orders drop out of review entirely — the first genuinely no-touch set.",
      "The medical necessity agent reads patient notes and the payer's policy criteria, queries a patient medical graph of extracted biomarkers, and returns an answered questionnaire with supportive and contradictory facts plus a confidence score; low-confidence or under-evidenced cases go to human escalation. The team has a publication on this work.",
      "The submission agent mirrors the EV/BV agent — per-payer customized integrations built from LLM-driven config and the shared tool repository — and all four agents have since been generalized to other patient workflows beyond prior auth."
    ],
    "takeaways": [
      "Start with deterministic checks and bring in agents only for the decisions rules can't make — don't lead with the LLM.",
      "Don't trust a single non-deterministic extraction as a basis for removing human review; find a second, independent source (prior auth letters, payer rule documents, portal checks) and act automatically only where the sources concur.",
      "Make confidence an explicit, attached output of the clinical reasoning step so escalation is a threshold decision rather than a blanket policy of reviewing everything.",
      "Treat browser automation as fragile by design: generate portal configs with an LLM over a shared action repository, and run a self-healing loop in production hours so breakages are mitigated rather than becoming failures.",
      "Look for whole units of work (here, entire orders) that become no-touch as a side effect of per-item classification — that's where the human-review reduction actually shows up."
    ],
    "topics": [
      "agents",
      "healthcare-ai",
      "prior-authorization",
      "rpa",
      "llm-extraction",
      "confidence-scoring",
      "human-in-the-loop",
      "knowledge-bases"
    ],
    "tools": [
      "Risa Labs",
      "LLMs (unnamed)",
      "RPA / browser automation",
      "SQL database (payer rule knowledge base)",
      "patient medical graph (biomarker graph)",
      "payer portals and APIs"
    ],
    "quotes": [
      {
        "text": "It might improve the efficiency, but it will not eliminate the human.",
        "at": "07:25",
        "url": "https://www.youtube.com/watch?v=_cVfz88_j7A&t=445s"
      },
      {
        "text": "So, now, instead of just having one source, I have two sources that give me the same information",
        "at": "08:08",
        "url": "https://www.youtube.com/watch?v=_cVfz88_j7A&t=488s"
      },
      {
        "text": "answers simple and complex clinical questions per patient and attaches confidence score to any answer. So, we escalate only the ones that actually need a clinician.",
        "at": "12:26",
        "url": "https://www.youtube.com/watch?v=_cVfz88_j7A&t=746s"
      },
      {
        "text": "We used multi-source of evidence to beat the single source of evidence to add confidence to our cases.",
        "at": "16:11",
        "url": "https://www.youtube.com/watch?v=_cVfz88_j7A&t=971s"
      }
    ],
    "words": 2797
  },
  "summary_url": "/#_cVfz88_j7A",
  "transcript": {
    "html": "/transcripts/_cVfz88_j7A.html",
    "txt": "/transcripts/_cVfz88_j7A.txt",
    "vtt": "/transcripts/_cVfz88_j7A.vtt"
  }
}