Realtime voice agents replace tier-1 IVR
“Hung up before the answer.”
Cause
The telco partner withdrew the test trunk on day 24 of a planned 40-call measurement and the pair was reassigned; 11 calls were measured, which is not enough to answer the hypothesis either way.
What was tried, what was hit, what it means
The run was designed to find the latency floor for a realtime voice agent handling tier-1 telco triage from Sydney: three vendor endpoints, one AU-hosted and two US-hosted, forty scripted calls each over a partner's SIP trunk, with turn latency and caller abandonment as the measures. The hypothesis was that at least one configuration would sit under 800ms median turn latency, which the partner's own research put as the threshold below which callers stop talking over the agent.
Eleven calls were completed on the AU-hosted endpoint before the partner's network team reclaimed the trunk for a production change and the re-booking fell into a change freeze. The eleven calls gave a median of 710ms, which looks like a pass and is not one: eleven calls from one endpoint on one afternoon is a tried-tier observation, logged as such, and no claim was made from it. The two US-hosted endpoints were never measured under load.
This is an abandoned entry. Nothing was disproved. The field's not-yet recommendation (r-voice-not-yet) rests on cost and on the accent-handling evidence from other signals, not on this run. The latency question is back in the pile with the harness ready and the scripts written; the missing input is a trunk.
Returned notes · what the next proposer sees
Stopped on 2026-07-15 with 11 of 120 planned calls measured. The harness, scripts and consent forms are in the lab repo and still work. The 710ms median from the AU endpoint is a tried-tier number, not a result; do not cite it. Before proposing again, secure the trunk in writing with a change-freeze exclusion, because that is what ended the first attempt.
The difference between “someone tried this before” and “someone tried this in June, here is what they hit, here is what would need to be true now.”
Lessons
- 01A partial measurement that looks like a pass is the most dangerous kind; log it at the tier it earned.
- 02Third-party infrastructure is a dependency with a change calendar; get the calendar before the trunk.
- 03Abandoned is a fine outcome and a cheap one; dressing it as a result would have cost more.
Record
Field
Voice and Vision Lineage intact. The reasoning that killed this stays attached to everything it touched.Experiment
Voice agent latency floor for AU telco Kill · Kill if no endpoint is under 1,000ms median after 20 calls each, or if caller-abandonment on the scripted calls exceeds 30% on every endpoint.Try again?
Proposals that touch this entry get its returned notes attached automatically.