Our Take
Most voice AI companies fail in emerging markets not because the technology is wrong, but because they optimized for GPU-rich Western infrastructure and standard English—leaving a real market opening for builders willing to work locally.
Why it matters
Voice remains the dominant customer service channel in Africa and the Middle East, where call volumes are three times higher than Western counterparts (per 4DX Ventures). Incumbents built for U.S./European workflows have no economic incentive to retrofit for dialects, code-switching, and telephony infrastructure that looks nothing like Silicon Valley.
Do this week
Infrastructure teams: audit whether your voice AI stack assumes Western GPU availability and English-only speech patterns—if so, test whether regional alternatives handle your actual call volume and latency targets.
AethexAI built voice AI from scratch for markets the big players skipped
Mariama Diallo (ex-Goldman Sachs, formerly at YC-backed ModelML) and Ayooluwa Odemuyiwa (ex-Meta, Caltech graduate) founded AethexAI last year to fill a gap: voice AI platforms that work in Africa and the Middle East. The company just closed a $3 million pre-seed round led by 4DX Ventures, with backing from Enza Capital, Dorm Room Fund, Mojo Ventures, Stanford GSB 26 Fund, and individual investors including Stanford faculty, telecom executives, and AI researchers from Anthropic (company-reported).
Rather than deploy existing orchestration tools like Vapi or LiveKit, the team built its own small models and orchestration layer. The Kora series ranges from 300 million to 1.7 billion parameters—a fraction of the size of standard LLMs. The company trained these models using anonymized call-center recordings and collected additional audio by shipping hard drives to radio stations across Africa. A network of university students annotated data and pronounced local names to keep costs down.
The result: AethexAI is now handling more than 17,000 calls per day (company-reported). The platform ships APIs and SDKs for developers and is offering onsite demos and workshops to help enterprises identify automation use cases. Current deployments focus on debt collection, customer activation, and KYC (Know Your Customer) verification.
The latency problem exposed a structural gap no incumbent had reason to solve
Call centers in Egypt that automated significant call volume had to roll back their systems due to poor results, according to the founders. Support centers across Africa reported that finding engineers to automate calls at acceptable cost was persistently difficult. The core issue: latency and jitter on automated calls in the region were "outrageous," in Odemuyiwa's words.
Most major voice AI providers (ElevenLabs, Deepgram, Sierra, Cognigy) were built for and optimized around Western markets. That means they assume high-end GPU infrastructure, standard English, and enterprise workflows common to the U.S. and Europe. When they enter emerging markets, they bring that same architecture—which creates latency, requires hosting models outside the region, and fails to handle dialects, code-switching, and informal speech patterns that dominate actual customer service in Africa and the Middle East.
Per 4DX Ventures co-founder Walter Baddoo, enterprises in Africa and the Middle East process roughly three times the call volume of their Western counterparts, as voice remains the dominant channel for customer interaction. Incumbent systems have neither the incentive nor the architecture to close this gap. AethexAI is betting that localized small models, on-the-ground partnerships with telcos, and infrastructure built for the region represent a sustainable opening.
Assess whether your voice AI vendor was built for your region
If you operate call centers in Africa, the Middle East, or any region where standard English and Western infrastructure don't dominate, test latency and accuracy on your actual dialects and call patterns before committing to a platform. Plug-and-play global solutions often fail in high-volume, low-latency environments outside North America and Europe. Diallo and Odemuyiwa advise customers to start with a single, high-impact use case (debt collection, activation, KYC) rather than attempting full automation, then expand once the system proves reliable in your specific market context. If your incumbent vendor cannot demo performance on your regional speech patterns or offer local infrastructure and support, treat regional alternatives as a serious option.