Phonely and Groq team up to kill the voice AI pause

Anyone who has ever called customer service knows the telltale sign: a pregnant four-second pause that immediately reveals you are talking to an AI.

Now, AI phone support startup Phonely, inference optimization platform Maitai, and chipmaker Groq have teamed up to eliminate that conversational uncanny valley.

By slashing response times by over 70% and pushing conversational accuracy from 81.5% to 99.2% across four iterations, the trio claims to have surpassed GPT-4o's 94.7% benchmark. Even more striking, over 70% of human callers reportedly cannot tell they are speaking to a machine.

Killing the four-second pause

Phonely CEO Will Bodewes pointed out that major frontier models suffer from massive latency variance. Roughly once every ten calls, the system freezes for up to four seconds. On a live phone line, that feels like an eternity.

To fix the stutter, Groq deployed what it calls zero-latency LoRA hotswapping. Instead of running a single monolithic model or taking a performance hit when swapping task-specific tweaks, Groq stores specialized LoRA adapters directly alongside base model weights in high-speed SRAM.

Sitting in the middle is Maitai, acting as a thin proxy orchestration layer. Maitai dynamically routes each request to the right specialized variant, pinpoints conversational weak spots, and incrementally fine-tunes models without human intervention.

At the 90th percentile, time-to-first-token plunged 73.4% from 661 milliseconds to just 176 milliseconds. Total completion time dropped from 1,446 milliseconds to 339 milliseconds.

Why it matters

For call center operations, this is not a distant prototype. One Phonely customer is already replacing 350 human agents this month alone.

If you run customer support, the barrier to voice AI adoption has officially collapsed. The primary complaint was never that AI lacked information; it was that robotic delays destroyed the customer experience. Sub-200ms latency effectively removes that friction.

For the AI industry, this partnership validates a growing architectural shift. Trying to solve real-time voice with a massive generalist model like GPT-4o looks increasingly like brute force. Specialized hardware running tightly orchestrated, hotswapped micro-models appears to be the smarter, cheaper path.

That said, we think reaching a 99% illusion rate carries hidden operational risks. When customers cannot tell they are speaking with software, regulatory and ethical disclosures become a minefield. Furthermore, the remaining 0.8% failure rate will consist of the hardest edge cases, where hallucinations during a convincing phone call can trigger massive customer backlash.