{
  "video": {
    "id": "qStB9GbppMU",
    "title": "1 Trillion Phone Calls/yr, 10% Error rate: The Crisis in Voice AI — Sumanyu Sharma, Hamming AI",
    "duration": 964,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "qStB9GbppMU",
    "title": "I Monitored Crime Audio. Voice Agents Scare Me More. — Sumanyu Sharma, Hamming AI",
    "one_liner": "Hamming's founder, who previously monitored thousands of hours of police radio at Citizen, argues voice agents now scare him more than crime: across the 10,000 agents Hamming monitors the real-world error rate is closer to 10% (not 1%), and their red-teaming product can break roughly one in five production agents.",
    "summary": "Sumanyu Sharma contrasts his old job listening to police scanners at Citizen with today's voice agent deployments: crime is decreasing, hyper-local and decentralized, while voice agents are taking off, centralized, and one prompt change has a blast radius across millions of users. He argues reliability is still the number one thing holding back production voice deployments — agents sound confident and natural while giving wrong information, skipping verification steps, or claiming they booked an appointment they didn't. He lays out a six-step debugging loop borrowed from ex-Facebook growth people (identify, prioritize by frequency and severity, understand, execute, check for regressions, monitor) and a maturity ladder from listening to calls manually, to rubric spreadsheets, to LLM-as-judge evals, to cross-conversation analysis. He closes on adversarial risk: Hamming's red-teaming product breaks about one in five agents, so his recommendation is now pre-deployment testing plus 24/7 red teaming.",
    "key_points": [
      "At least a trillion calls happen per year and most will be handled by conversational voice agents within five years; at a 1% error rate that's 10 billion incidents a year — but Hamming monitors 10,000 agents and sees an error rate closer to 10% in practice.",
      "Observed failure modes: agents claiming they found the right policy after skipping eligibility or verification steps, applying discounts they shouldn't, mishearing the user, giving incorrect information, and claiming an appointment was booked when it wasn't — which happened to the speaker personally, costing him two wasted hours at a physician's office.",
      "Crime is hyper-local and decentralized (a robbery affects a finite set of people); voice agents are centralized, so a single prompt or architecture change has a massive downstream blast radius across millions of users.",
      "Severity spans from 'trash fires' (repetition, not quite understanding the user — annoying but harmless) to genuine safety risks, e.g. a drive-thru agent at a Taco Bell or McDonald's handling a vegan burger order with peanut allergies.",
      "Six-step loop from ex-Facebook growth friends: identify problems, prioritize by impact/frequency/severity, understand the fix, execute, check whether the change worked and caused regressions, then keep monitoring in production. The fix itself is the lowest-effort part of the pipeline.",
      "Two axes for coverage: known problems (turn latency, interruptions, ASR issues) tracked over time, versus emerging behaviour only visible across many conversations. Teams progress from manual listening → a spreadsheet of 5–10 rubrics (greetings, closing, validation, core logic) → LLM-as-judge eval products → cross-conversation analysis, which is where the best teams now spend time.",
      "Priority matrix: one-off + low impact = who cares; systematic + low impact = still worth fixing, especially in a bake-off; systematic + high impact = P0, e.g. a fintech agent that fails to freeze a customer's credit card.",
      "Hamming shipped a red-teaming product in April and can break roughly one in five agents across financial services, healthcare and consumer — bypassing verification, prompt-injecting agents, and extracting data they should not have. 'This is not theoretical.'"
    ],
    "takeaways": [
      "Start by listening to real calls manually — don't skip it, it builds intuition — but move up to cross-conversation analysis, because per-call evals only check consistency on known problems and never surface novel emerging patterns.",
      "When validating a fix, don't just replay the exact failed conversation 5/10/20/50 times to measure pass probability; keep the intent but vary wording, patterns, accents, style and add extra intents, so you know your change is net positive rather than net negative.",
      "Use A/B testing in production for hypotheses that synthetic pre-deployment testing can't settle — e.g. the first 5 seconds of an outbound call, where vocal quality and exact wording matter most.",
      "Run 24/7 red teaming against your own agents, especially where bad interactions are costly; every capability, tool and data source you add to make the agent more useful widens its attack surface.",
      "Monitor both sides of the call: not only what the agent says and does, but whether users are being adversarial and trying to trick it into doing things it shouldn't."
    ],
    "topics": [
      "voice-agents",
      "reliability",
      "evals",
      "red-teaming",
      "ai-safety",
      "monitoring",
      "production-ai",
      "prompt-injection"
    ],
    "tools": [
      "Hamming",
      "Citizen",
      "Aqua Voice",
      "Super Whisper",
      "Whisper Flow",
      "LLM-as-a-judge",
      "Taco Bell",
      "McDonald's"
    ],
    "quotes": [
      {
        "text": "so I think voices sound very confident. They sound very natural, but the information provided is often, you know, not correct. That's the biggest problem here.",
        "at": "03:37",
        "url": "https://www.youtube.com/watch?v=qStB9GbppMU&t=217s"
      },
      {
        "text": "In practice, we currently monitor 10,000 agents and the error rate is closer to 10% in practice.",
        "at": "04:58",
        "url": "https://www.youtube.com/watch?v=qStB9GbppMU&t=298s"
      },
      {
        "text": "we can probably break one in five agents at this point. We've tested this across financial services, healthcare, um consumer and so on. We've bypassed verification.",
        "at": "13:10",
        "url": "https://www.youtube.com/watch?v=qStB9GbppMU&t=790s"
      },
      {
        "text": "the reason I fear voice agents more than crime is that one of these incidents is going to impact you. It already did for me.",
        "at": "15:16",
        "url": "https://www.youtube.com/watch?v=qStB9GbppMU&t=916s"
      }
    ],
    "words": 2967
  },
  "summary_url": "/#qStB9GbppMU",
  "transcript": {
    "html": "/transcripts/qStB9GbppMU.html",
    "txt": "/transcripts/qStB9GbppMU.txt",
    "vtt": "/transcripts/qStB9GbppMU.vtt"
  }
}