11x details latency and turn-taking changes in its AI voice agents
An 11x engineering post described changes to its sales voice agents, including speculative reply drafting and a split-voice architecture.
By Cora Whitley, GTM AI Analyst, Chief of Staff
Image: 11x press materials, from the announcement
11x published an engineering post on October 21, 2026 describing changes to the timing of the AI voice agents it runs on sales calls. The company says its original agents took about 950 milliseconds to respond, roughly five times slower than natural human pace. It now targets replies between 300 and 700 milliseconds and measures success as 95% of replies delivered under 700 milliseconds. The post was written by Francisco Izaguirre, a member of technical staff.
The details
- 11x says its text-to-speech provider closed the connection after every reply, and it now keeps a spare connection open so the handshake happens while the prospect is speaking.
- Agents draft replies while the prospect is still talking and discard them if the prospect continues, and drafts never commit irreversible actions such as booking a meeting.
- Agents say a filler line such as a request to wait a moment before looking up an answer, and a low room-tone track plays under the call.
- In the split-voice design, the main model hears the audio directly while a separate voice model produces speech, which lets customers choose voice and pronunciation.
- Parallel fast models act as co-drivers for email capture, pronunciation, safety guardrails and background-noise detection.
Context
The post cites published research that speakers in conversation typically respond within about 200 milliseconds, and an English phone-call study in which nearly a third of rejections came after 700 milliseconds. 11x also says it changed how it measures latency, since its earlier metric started only after the agent decided the prospect had finished speaking. The company sells inbound and outbound agents named Julian and Alice. The post lists matching a prospect's pace, choosing a voice per prospect and handling objections as further work. On interruptions versus delay, it states "Doing both is worse than being slightly slow."
Not disclosed
- Rollout dates for the changes
- Which products or plans include them
- Measured results against the 95% target
- Vendor
- 11x
- Category
- AI Agents & AI SDRs
- Announcement type
- Feature update
Report an error
Vendors and readers can flag a factual error in this story, or ask for an image or logo to be removed. Email Cora, the editor, with the link. Image and logo removals happen within one business day. For factual corrections, Cora replies with next steps and a ForgeX lead approves the fix before it is published.
