AI Outreach Personalisation: The 6-Rung Ladder
Personalised outreach at scale has six rungs, defined by what each message is conditioned on. The conditioning ladder, the break-even maths, and the crossover.
Personalised outreach at scale has six rungs, defined by what each message is conditioned on. The conditioning ladder, the break-even maths, and the crossover.
Most AI voice agents do not need a SIP trunk. The crossover volume, worked from published Retell, LiveKit, Twilio and Telnyx rates, and the five exceptions.
Voicemail detection in AI voice agents fails five documented ways. Match your symptom to the layer that owns it, run the test, then fix the right threshold.
AI voice agents interrupt because a fixed silence timer, not meaning, ends the turn. Endpointing defaults from LiveKit, Pipecat, Vapi, Retell and Deepgram Flux.
Spelled email addresses fail because per-token errors compound: at 96.6% accuracy a 22-token address survives just 46.7% of the time. The maths and the fixes.
A spoken card number pulls four stores into PCI DSS scope: transcript, model context, platform logs, call recording. The component table, and the safer route.
What happens when my AI agent hits a phone tree: the six stages of the call, why classification is slowest, DTMF transport, and a 3/90 rule for stopping.
ElevenLabs publishes 75ms, Deepgram 80ms, Cartesia sub-90ms, Rime 37ms. We traced each to the vendor page on 10 September 2026 and recorded what it excludes.
Ten minutes of speech-to-speech costs $6.90 in tokens when caching misses, $0.93 when it holds and $0.22 as a cascade. Worked from vendor pricing pages.
Your AI dialler called a customer at 9pm their time. Here is the 24 hour triage, the safe mainland calling window in UTC, and the timezone fix that stops it.