Phonely has released Alma, a bespoke language model engineered specifically for voice-based customer interactions rather than adapted from text-trained models. The distinction matters operationally: most contact centres currently rely on GPT-4.1, a model OpenAI has deprecated, because frontier labs have since optimised for long-running agents and coding capabilities at the expense of latency and conversational naturalness. Alma inverts this priority by training on how people actually speak rather than how they write, which Phonely claims produces faster response times, lower costs, and superior conversion rates. The model achieves this through a hybrid architecture deploying four specialised LLMs in sequence—one handling task verification, another managing conversation flow, a third extracting information, and a fourth orchestrating the process—which increases determinism without claiming perfect reliability. This architectural approach directly addresses the latency spikes contact centres experience with OpenAI's shared infrastructure, where requests queue unpredictably and can delay responses by 30 seconds or more during peak usage.
For CX teams evaluating voice AI vendors, Alma's release signals a meaningful shift in how the market is segmenting. Rather than competing on general-purpose capability, Phonely has built a narrower, purpose-built tool that trades breadth for depth in voice-specific performance. This raises a critical question: as specialised models proliferate, will your organisation benefit from best-of-breed voice solutions, or does the operational overhead of managing multiple vendors outweigh the performance gains? The answer depends partly on your current infrastructure—teams already running Agentforce or similar enterprise platforms may find integration friction outweighs Alma's latency advantages, whilst mid-market operations handling high-volume inbound calls will likely see immediate ROI from reduced response delays and improved first-contact resolution.
The broader implication concerns how contact centres should evaluate voice AI readiness. Bodewes emphasises that voice agents require deliberate training and iterative refinement, much like onboarding human staff, yet many organisations still approach AI deployment as a plug-and-play capability. This gap between expectation and reality means your vendor selection should prioritise not the model's raw capability but the tooling available for training, monitoring, and continuous improvement. Teams that treat voice AI as a hire-and-coach function rather than a one-time implementation will extract substantially more value, regardless of which underlying model powers the agent.
Businesses need voice-based AI agents that can interact naturally, quickly and follow established processes. With Alma, Phonely seeks to meet those needs.