Why Now
Four shifts make real-time agentic AI shippable to small businesses — for the first time.
Real-time agentic orchestration over voice has been technically possible for years. It became economically and ergonomically viable in the last 18 months. The window is open and incumbents have not moved.
1. The latency cliff is now under a second
Streaming ASR, streaming LLMs, and streaming TTS finally compose end-to-end. A voice agent can hold a conversation at sub-second response time without exotic infrastructure. Below ~1s, callers stop noticing they're talking to software.
2. GPU economics for real-time inference
Per-token inference cost on frontier models is down an order of magnitude in two years. Per-call voice economics now work at trade-services pricing — what used to be a research project is now a unit-economics question.
3. APIs are standardized — orchestration is the moat
Every business system speaks REST, OAuth, webhooks. Integration is no longer the blocker. The hard problem moved up the stack: holding state across channels, gating tool calls, surviving telephony edge cases. That's where engineering matters now.
4. Service trades remain underserved
Painters, cleaners, and small remodelers are 6M+ US businesses running on phones, paper, and group texts. Vertical SaaS in this segment has chosen scheduling and dispatch — nobody has shipped a voice front desk that actually works on the phone.
What didn't change
The hard parts are still hard.
We're not bullish on every voice-agent claim. A lot of demos work in a quiet room and break on a phone line. The problems we still respect:
Echo & barge-in
Single-channel telephony bleeds the agent's audio back into the caller's track. Without DSP-level echo cancellation and proper VAD, the agent talks over itself or refuses to interrupt. Most demos skip this.
Tool honesty
An LLM that confidently says "I booked your appointment" without calling the tool is worse than no agent at all. State-machine validation isn't optional — it's the difference between a product and a liability.
Telephony compliance
10DLC, A2P consent, recording disclosure, per-tenant brand registration. Carriers will shut down a line that ignores these. Solo founders learn this the hard way.
Vertical vocabulary
"Sherwin Emerald in eggshell," "240V receptacle," "PEX vs. copper." Generic ASR loses these. Trade-tuned models are the cleanest case for self-hosted inference once volume justifies it.
The window is open. We're walking through it.
Frontbell opens its private founder-cohort beta in June 2026; public launch is mid-July 2026 for painters, cleaners, and small remodelers. The engines underneath were designed to bring the same product to additional trades after launch.
See the product →