Investors have spent heavily on voice AI, from customer-service platforms to meeting notes and dictation. On 11 October 2026, the people building it told TechCrunch the obvious part is done and the hard part is not. The voices can sound human. The systems still miss words, answer too slowly, and lose trust the moment a follow-up action is wrong.
What the operators said
PolyAI CTO Shawn Wen said full-duplex models, which can speak while they listen, are a milestone, and that the next problem is reasoning fast enough for the conversation to feel natural. He wants a caller to stay for two or three turns and believe the agent can solve the problem. Otter CMO Alex Gay said speaker identification, intent, and company knowledge are what turn a transcript into automation. He also said a bad transcript poisons every action after it. Both said people should know when they are talking to AI, or when a meeting is being recorded. TechCrunch: https://techcrunch.com/2026/10/11/these-execs-think-voice-ai-hasnt-reached-its-chatgpt-moment-yet/
What that means on a real phone line
- Speed is part of the product. A correct answer that arrives after an awkward pause feels broken.
- Keywords pay the bills. A missed product name, PIN code, or locality sends the wrong task to the CRM.
- Language is not solved. Indian English, Hindi, and mixed calls need their own test set. A US demo will not catch them.
- Say it is AI. A caller who discovers the agent later will not trust the next call.
This matches how we compare a menu with a conversational agent in AI voice agent vs IVR. Keep the menu for three clear routes. Use an agent where the caller describes a job in their own words, and only after the transcript is good enough to act on.
A pilot that respects the gap
- Pick one call reason, such as “book a site visit,” and write the fields the agent must capture.
- Run 30 real recordings, including noisy lines and mixed language. Count missed names, numbers, and places.
- Block any action that depends on a field the transcript got wrong.
- Tell the caller they are speaking with an assistant, and transfer to a person on request.
FAQ
If full-duplex models exist, why wait?
Talking and listening at once is not the same as fetching the right answer from your stock, your calendar, or your policy. Wait until that lookup is fast and checked.
Are meeting notes safer than phone agents?
They fail in the same place. A wrong summary that creates a task is still a wrong action. Read how businesses use AI calling agents before you let notes write into a CRM.
Test the words before the voice
Contact Altron Technologies with a sample of real calls. We will tell you whether an agent or a shorter menu fits.