Three things converged in 2025 and 2026. (1) Voice traffic from customers is up across every segment we have data on. (2) The cost of running our own carrier-grade SIP infrastructure fell off a cliff. (3) AI voices finally cleared the uncanny valley and TTS latency dropped under 200ms.
Why voice is up
Our own data: 38% of conversations in Q1 2026 included at least one phone leg. That number was 22% in Q1 2025. The pattern is consistent across SaaS, ecommerce, healthcare, and real estate customers.
Theories: customers got tired of chatbots that could not deliver, AI did not eat phones the way pundits predicted, regulated industries cannot SMS without consent friction, and customers in their 50s and 60s are more comfortable calling than chatting. Pick the one that fits your customer base, all of them are part of the answer.
Why we self-host
“We replaced Aircall in 28 of our last 30 onboardings. The 2 holdouts had legacy Cisco gear.”
Aircall is $30 per user. OpenPhone is $20. Dialpad is $23. At a 28-person team, that is $560 to $840 a month for phones alone. We run our own carrier-grade SIP infrastructure at roughly $0.005 per inbound minute and $0.015 per outbound. A 28-person team running 100k minutes per month pays roughly $1k in usage. The savings are not 100%, but the freedom to do AI on the call (which Aircall does not allow) is.
Where the AI fits
Two places. First, post-call: every recording is transcribed on our own infrastructure, summarized, and pinned to the conversation. Second, on-call: when the caller is on hold, the AI engages with your brand voice, grounds in your KB, and offers to text the answer. We do not let the AI run free on the call itself yet (legal risk, customer expectations), but the wait-time concierge alone has cut average call time by 31 seconds across customers running it.
What is next
Phase 2: the AI shadows live calls and feeds the agent suggestions in real-time. We have a working prototype. The agent sees a transcript with the suggested next reply as the customer is talking. It is uncanny how much it shaves off agent training time.