{
  "video": {
    "id": "-a2oTS9J67U",
    "title": "The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale",
    "duration": 905,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "-a2oTS9J67U",
    "title": "The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale",
    "one_liner": "Getting an agent to work is only the first step. Before it can run at scale, leadership has to answer three questions: what value it creates, who is accountable for it, and what evidence lets it run with more autonomy.",
    "summary": "Michele Grant draws on 20+ years of enterprise transformation and more than a decade of M&A integrations. She argues that trust in an AI agent has to be engineered before day one rather than hoped for after launch, and that deciding each agent's permissions, owner, evidence and escalation triggers is an executive operating mandate, not an engineering task. Her worked example is a $15B acquisition targeting $400M in net synergies across 40 workstreams, where an agent drafts the weekly steering-committee synergy report and matched analysts line for line in a four-week pilot. She covers the value case and its baselines, accountability under the three lines model and a 'human work standard', and three engineering artifacts that earn the agent autonomy over four stages.",
    "key_points": [
      "A pilot tests whether the software functions. Daily operations test whether it creates value within the controls leadership has set. Risk frameworks supply principles but not how to apply them per agent, so for every agent leadership must decide what it may do, who owns it, what evidence it records and what triggers an escalation.",
      "The value case is written and agreed before the agent touches production data. It has three metrics: cycle time (e.g. 5 business days down to 2), analyst capacity (4 analysts spending half their week = 80 hours/week, counted only when the integration lead reassigns that time to lagging workstreams, and net of compute, licensing and human review cost), and detection time, which she calls the most important. A faster report does not guarantee slips are seen sooner.",
      "Every metric needs a pre-deployment baseline: 'If the IMO cannot state the baseline today, do not deploy.'",
      "Under the three lines model, management owns the controls (the integration lead is the accountable executive), risk and compliance oversee and challenge in proportion to the agent's exposure, and internal audit decides when the agent enters the annual audit plan. One named daily overseer (the senior analyst) must understand the agent's boundaries, interpret its output in context, have authority to override or reverse it, and have a clear kill switch.",
      "A one-page 'human work standard' maps every decision fork to the human role authorized to make it, the required evidence and the escalation path. The agent prepares the analysis and stops there. The only reason to delay deployment is a decision with no assigned owner.",
      "Evidence artifact 1 is an auditable decision log: raw data ingested, tool calls, prompt, model versions, output and human sign-off. Core math and reconciliation run in deterministic, finance-approved code, not LLM prose. The model's summary is an explanation, not proof; the real proof lives in the telemetry. If reported savings don't match the general ledger, the agent doesn't touch the number but flags the discrepancy and routes it up.",
      "Evidence artifact 2 is a model inventory record: accountable executive, data lineage, source-system dependencies, validation results, escalation path, kill-switch runbook, and regression triggers (base model or source API updates that require full revalidation).",
      "Evidence artifact 3 is an autonomy scorecard with four stages: shadow mode, supervised draft, exception and sample (lagging initiatives plus a random sample of claims), and exceptions only with periodic independent sampling. It puts control metrics beside value metrics so the steering committee can promote the agent, hold it or step it back. A falling override rate may mean review depth has slipped, not that the agent got smarter."
    ],
    "takeaways": [
      "Write down baselines for cycle time, capacity and detection time before deploying, and don't deploy if you can't state them. Count freed hours as value only once they are reassigned, and net out compute, licensing and review costs.",
      "For every agent, name an accountable executive and a daily overseer who has override authority and a kill switch. Map each decision fork to a human role in a one-page human work standard.",
      "Run core calculations and reconciliation in deterministic, approved code. Log raw inputs, tool calls, prompts, model versions, outputs and sign-offs, and treat LLM summaries as explanation, not proof.",
      "Keep a model inventory record with regression triggers, so any base-model or source-API change forces full revalidation before the agent runs in production again.",
      "Grant autonomy in stages (shadow, supervised, exception and sample, exceptions only). Once humans review only exceptions, use independent sampling to catch what the agent failed to flag and to read override-rate trends correctly."
    ],
    "topics": [
      "agents",
      "ai governance",
      "trust",
      "risk management",
      "enterprise adoption",
      "human in the loop",
      "observability",
      "m&a integration"
    ],
    "tools": [
      "Leading at Scale"
    ],
    "quotes": [
      {
        "text": "Go-live proves that a system exists, but it doesn't prove the people behind the work will actually use it.",
        "at": "00:02",
        "url": "https://www.youtube.com/watch?v=-a2oTS9J67U&t=2s"
      },
      {
        "text": "If the IMO cannot state the baseline today, do not deploy.",
        "at": "06:17",
        "url": "https://www.youtube.com/watch?v=-a2oTS9J67U&t=377s"
      },
      {
        "text": "A falling override rate does not automatically mean the agent is getting smarter. It can also mean the depth of review has slipped.",
        "at": "13:14",
        "url": "https://www.youtube.com/watch?v=-a2oTS9J67U&t=794s"
      },
      {
        "text": "Building an agent that works is just the first step.",
        "at": "14:48",
        "url": "https://www.youtube.com/watch?v=-a2oTS9J67U&t=888s"
      }
    ],
    "words": 2056
  },
  "summary_url": "/#-a2oTS9J67U",
  "transcript": {
    "html": "/transcripts/-a2oTS9J67U.html",
    "txt": "/transcripts/-a2oTS9J67U.txt",
    "vtt": "/transcripts/-a2oTS9J67U.vtt"
  }
}