I Monitored Crime Audio. Voice Agents Scare Me More. — Sumanyu Sharma, Hamming AI

Sumanyu Sharma showed up for a doctor's appointment a voice agent had told him was booked. He was not on the schedule, the front desk turned him away, and he lost two hours. His point is not the inconvenience but the substitution: make that his grandparent, and make it a procedure rather than a checkup, and the same failure costs something else entirely. Sharma founded Hamming after years at a public safety app, where he and his team listened to thousands of hours of police radio and pushed millions of alerts across several US cities. He puts the two experiences side by side deliberately. Crime is decreasing and it is local, touching whoever happens to be involved. Voice agents are scaling fast and they are centralized, so one prompt change propagates to everyone at once. Around a trillion phone calls happen every year. Even at a one percent error rate that is ten billion bad interactions, and across the ten thousand agents his company monitors the real rate is nearer ten percent. The failures are rarely dramatic. An agent reports it found the right policy while quietly skipping the eligibility check. It applies a discount nobody authorized. It says it booked something it did not. His remedy is a loop rather than a fix: find the problems, size them by frequency and severity, change something, verify the change did not break something else, and keep watching. He is emphatic that listening to individual calls by hand is where to start and not what to scale, and that the real insight lives in patterns across conversations. Then the harder warning. His team's adversarial testing breaks roughly one agent in five. Speaker info: - https://x.com/sumanyu - https://www.linkedin.com/in/sumanyusharma/ - https://hamming.ai Timestamps: 0:00 - Listening to crime audio at scale 2:50 - Why reliability still blocks deployment 4:32 - Crime is local, voice agents are centralized 5:23 - Sizing severity, from annoying to unsafe 7:04 - A loop for finding and fixing failures 7:57 - Coverage: from listening by hand to cross call analysis 10:32 - Testing whether a fix really worked 12:13 - When bad actors learn to dial 13:07 - Breaking one agent in five