Voice agents crossed a threshold recently: the latency is low enough and the voices natural enough that callers stop noticing they're talking to software. That changes the math for any business that lives on the phone.
Where a voice agent earns its keep
Not everywhere. The sweet spot is high-volume, low-variance calls: opening hours, order status, booking changes, first-line triage. If 60% of your calls are the same ten questions, those calls are automatable today.
The parts you actually need
- Speech-to-text that handles real phone audio, accents and all.
- A language model grounded in your data — your FAQ, your prices, your booking system. An agent that improvises answers is a liability.
- Text-to-speech that sounds like a person, not a satnav.
- An escalation path. The fastest way to lose a customer is trapping them in a bot loop. Ambiguous or emotional calls must reach a human, with context attached.
What it costs
Less than you'd think, and the economics keep improving. Self-hosted open models mean no per-minute platform fees for the core pipeline — you mostly pay for a modest server and the integration work.
The honest caveats
Voice agents fail badly when they're deployed as a cost-cutting wall between customers and humans. They succeed when they handle the boring 60% so your people can be excellent at the interesting 40%.
Thinking about this for your own business? Get in touch — I've built this stack end-to-end.