3 months running an AI voice agent in a real business: what worked and what didn't
Admir-Rusidovic · reddit · 2026-10-17
A business owner shares lessons from 3 months of running a self-built AI voice agent on live calls, with plans to turn it into a SaaS:
- Capturing basics was hardest: no STT model reliably handled postcodes, names, emails. Fixes: knowledge-base fuzzy matching for postcodes/names; for emails, an SMS review link after calls (60% completion)
- Latency: they once masked silence with keyboard-typing sounds ("straight from the 1980s") before fixing background processing
- Noise: street chatter caused hallucinated words; noise cancellation plus context filtering largely solved it
- Customers liked it: most callers were fine, some asked what models it runs or tried to get cake recipes; the agent always discloses it's AI — the author rejects human-mimicry
- Biggest win isn't cost: it's uninterrupted focus — review collected info afterwards and decide what needs a callback
More from coding & agent
- CoPE adds one forecasting loss to GRPO, doubling long-horizon agent sample efficiency — Lianhuiq · 2026-10-17
- laya-guard 0.5.0 adds AI attack simulation and file-tool hooks to guard Claude Code — Extreme_Ad5709 · 2026-10-17
- Laude's open-source RevEngine review bot takes #1 on GitHub ReviewBench in a week — lateinteraction · 2026-10-17
- Laude open-sources Headlong: a sub-10K-line Bash microharness for agents that never stop thinking — lateinteraction · 2026-10-17
- Hacktoberfest 2026 pivots to open-weight AI: 500+ events, 3,000 projects in week one — jonmarkgo · 2026-10-17
- MCP server runs Claude Code, Codex and Gemini as subagents with account rotation — Kynori20 · 2026-10-17