The voice agent that answers the phone, and the voice that lies on it
One credit union's agent handled 598,000 calls in 11 months. The same week, prosecutors indicted a $129M scheme built on cloned CEO voices. The phone is now an AI battleground.
- SoundHound reports American Heritage Credit Union's AI agent Gigi handled 598,000 member conversations in 11 months and resolved 88% of fraud and dispute chats without a human representative from July 2025 through May 2026. Self-reported m:…5203
- Cekura reports PolyAI ranked first among ten voice agent platforms in its Bench workflow benchmark, passing 87.8% of 82 caller scenarios on all three runs, with a pass³ consistency score 12.2 points ahead of the second-place platform. Self-reported m:…7667
- Liquid AI says models trained on LiquidCodec tokens achieved the highest naturalness scores across eight test sets in zero-shot multilingual text-to-speech, and the codec encodes at 361x real-time on an Apple M4 Max. Self-reported m:…9241
- Resemble AI reports federal prosecutors indicted three individuals for a $129 million fraud scheme that used AI-cloned CEO voices and fake outside counsel to authorise wire transfers. Self-reported m:…3559
- Reality Defender says agentic AI allows a single person to launch thousands of identical threats simultaneously from 10,000 miles away, and the company is offering its voice detection technology free to schools, 911 dispatch centres and emergency responders. Self-reported m:…5109
The agent answers
The voice agent has stopped being a demo. SoundHound reports that American Heritage Credit Union's AI agent, Gigi, handled 598,000 member conversations in 11 months — about 1,897 phone interactions a day — and resolved 88 percent of fraud and dispute chats without a human representative between July 2025 and May 2026. The agent handles more than 30 types of member requests around the clock, and the credit union is now adding agentic workflows so it can handle more complex requests on its own. A credit union with more than 5 billion dollars in assets has put its members' fraud calls in the hands of a voice agent, and the numbers say it is working. m:…5203
The benchmarks are catching up to the deployment. Cekura reports that PolyAI ranked first among ten voice agent platforms in its Bench workflow benchmark, which tests agents on complex real-world scenarios such as clinic appointment scheduling and Medicare intake rather than isolated model performance. PolyAI passed 87.8 percent of the benchmark's 82 caller scenarios on all three runs, completed the caller's task on 98.8 percent of calls, and led the second-place platform by 12.2 points on the pass-three consistency metric. PolyAI's own account of the shift is that the product is no longer a voice agent but a dialog agent: one that completes work in systems of record, built on an audio-native model, Dialog-RSN-1, that handles turn-taking, speech recognition and function calling in a single model. The test that matters is no longer whether the voice sounds human. It is whether the job gets done, three times in a row. m:…7667 m:…6457
The stack gets faster
Underneath the agents, the audio plumbing is getting faster. Liquid AI has released LiquidCodec, an in-house streaming audio codec built to support its LFM3-Audio-2B and LFM3-Omni-2B models. It uses grouped finite scalar quantisation to allow parallel prediction of audio tokens, and the company says models trained on its tokens achieved the highest naturalness scores across eight test sets in zero-shot multilingual text-to-speech. On an Apple M4 Max, the codec encodes at 361 times real-time and decodes at 87 times real-time. Voice agents are no longer limited by how fast they can turn speech into tokens; the constraint has moved to what the agent does with them. m:…9241
The same voice, weaponised
The same week, the other side of the voice story is a police blotter. Resemble AI reports that federal prosecutors indicted three individuals for a 129 million dollar fraud scheme that used AI-cloned CEO voices and fake outside counsel to authorise wire transfers, targeting thousands of businesses between July 2024 and February 2026. The same report notes Russian military units allegedly using AI-generated drone footage of flag-raisings to deceive their own chain of command, with at least 35 suspect videos identified by the Institute for the Study of War, and the estates of Elvis Presley and Eminem's publisher suing Suno over voice likeness. Fourteen percent of public safety communications professionals surveyed by APCO reported being targeted by AI tools. The cloned voice is now a fraud instrument, a propaganda instrument and a legal fight, all at once. m:…3559
The most urgent version of the threat is the cheapest. Reality Defender says AI-generated voice threats are increasingly used to trigger school evacuations and disrupt emergency services, and that agentic AI allows a single person to launch thousands of convincing, automated threat calls simultaneously from anywhere in the world — the company puts it as thousands of identical threats from 10,000 miles away. The company argues existing laws are ineffective against anonymous, overseas perpetrators, and is offering its real-time synthetic voice detection technology free to schools, 911 dispatch centres and emergency responders, while Orange Business integrates it into its Live Intelligence platform and Samsung Electronics Singapore has signed an agreement to deploy on-device deepfake detection to Galaxy devices. The phone call is now two products at once: the agent that resolves the call, and the detector that verifies the voice on it. Both are shipping this week. m:…5109
What is still unresolved
- Whether voice-agent resolution rates hold as agents take on more complex requests
- Whether on-device deepfake detection can keep up with cloned-voice quality
- Whether free detection for schools and dispatch centres becomes standard infrastructure
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.