← News

ResearchSources: arXiv · 16 Nov 2026

FEND classifies who answers outbound Voice AI calls: human, voicemail, IVR or another AI

Researchers behind the FEND framework say standard answering-machine detection is too coarse for AI agents making calls, and propose per-class detectors that label the far end within 20 seconds.

In plain words

When an AI agent phones someone, the call may be picked up by a person, a voicemail system, a menu-driven phone system (IVR) or another AI. The authors argue a simple human-versus-machine signal cannot tell the agent which of these very different situations it faces. FEND instead runs separate detectors for each automated type and returns one clear label.

What was announced

The paper reports that in 1,000 randomly sampled calls from the authors' own outbound deployment, 48.1% of far ends were live people, 44.6% voicemail systems, 6.9% other AI systems and 0.4% IVR systems. FEND reads the far end's audio and streaming transcript and runs one detector per automated class, with human as the default when nothing fires, returning a single label within 20 seconds.

On a public benchmark, the authors say the voicemail components reach 100% accuracy without misclassifying any live human, where Asterisk's AMD misclassifies 30 to 66% of them. For AI callees, acoustic features did not transfer across recording conditions, but speech timing did, separating AI from human at 0.952 ROC-AUC because synthesised speech leaves shorter gaps between words. On production traffic, separating a live person from any automated far end reached F1 0.979, with 2.9% of live people given an automated label, the error the authors call most consequential for an agent acting on the result.

What it does not show

  • Evidence is limited to the paper's abstract; the full article's methods, datasets and caveats were not reviewed.
  • Reported accuracy and ROC-AUC figures come from the authors' own evaluation and deployment, not from independent confirmation.
  • The AI-callee timing signal is described as sensitive to recording conditions, since acoustic features did not transfer across them.

For security and IT teams

Scope
The findings cover outbound Voice AI calls and classify far ends as human, voicemail, IVR or AI callee, based on the authors' own deployment and a public benchmark.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.