An AI dialer that books meetings, not a chatbot that wastes them

Half the AI-voice demos floating around look the same: a synthetic voice reads a polished script, a human nods on cue, and a marketing line about “AI agents that close deals” goes over the top. On an actual call with an actual prospect, that polished script falls apart inside thirty seconds.

What follows is what stuck after wiring our AI dialer to real prospects across thousands of calls.

Latency is the entire game

If the AI takes more than ~700 ms to start responding, the call feels broken. The prospect interrupts, the AI gets confused, the prospect hangs up. The clock has to be measured end-to-end:

1. Audio leaves the prospect’s mouth.
2. Speech-to-text catches a sentence boundary.
3. The LLM generates a response.
4. Text-to-speech starts producing the first phoneme.
5. That phoneme reaches the prospect’s handset.

Each step costs 50-200 ms, and three of them are network-bound, which means a single bad peering hop can quietly add a quarter of a second. Most of our infrastructure work goes into making sure none of those steps stalls long enough for the human to notice.

Scripts that recover

A good outbound script has three layers:

• Goal: the single outcome the call exists for (book a 20-minute meeting).
• Branches: the most likely four to six paths the conversation takes.
• Recovery: explicit graceful exits when the AI does not understand.

Treating the script like code - with regression tests that replay recorded calls and fail loudly when the recovery layer is missed - is what makes the dialer reliable. Picking a fancier LLM rarely is.

Hand-off, every time

The dialer always offers a path to a human. “Would you like a quick transfer to the team?” is the single most important sentence in the script. Customers who feel cornered by an AI do not call back, and the cost of an avoidable bad first impression is much higher than the cost of one extra transfer.