Synopsis: A caller waiting for an answer has little patience for a model working through a lengthy reasoning process. Voice AI has to respond quickly, understand the situation and connect to the systems needed to complete a task. Those requirements can favor smaller, specialized models over larger general-purpose ones. The model decision is as much about response time and the job at hand as it is about raw capability.
Will Bodewes, co-founder and CEO of Phonely, explores those trade-offs with Mike Vizard. Open-weight models give businesses more control over behavior and customization, but training and maintaining them requires expertise that can be difficult to recruit. Bodewes describes an alternative to building everything internally or relying entirely on frontier models: working with providers that fine-tune models for a particular task. Specialization can change the economics, although it does not eliminate the work of integrating the system into a business.
A voice agent still needs access to relevant data, tools and other services. Bodewes discusses orchestration and MCP connections as ways to make those resources available, while emphasizing the response-time constraints of a live call. Unlike a fixed-menu phone system, an AI agent can use context to respond dynamically. That flexibility needs training and testing, not an assumption that a convincing demo means the agent is ready for every customer interaction. Businesses should expect to refine its behavior as they learn where it succeeds and where it struggles.
Bodewes also challenges the habit of evaluating voice AI only as a way to reduce call-handling costs. Answering inquiries around the clock or converting more leads into appointments can create value beyond doing existing work more cheaply. Those outcomes give teams something concrete to measure as they improve the system. He distinguishes useful inbound assistance and requested reminders from unsolicited AI cold calls, which he opposes. The case for deployment rests on serving customers and improving results, not simply automating more phone traffic.
