Skip to research
SAVED EVIDENCEEditorial preparation

How a collections agent's style changes a voice call

Three fictional borrowers took calls from two versions of a reference collections agent. The script-first version drew more cut-ins than the listen-first version, 7 against 3, in six short calls that show the method rather than a general result.

Fieldwork
3 November 2026
Study scope
Six text calls of up to 8 turns: three virtual borrowers, two agent versions. Voiced in Hinglish, with an English re-voicing, for 12 call versions.

THE SAVED OBSERVATIONS

What the evidence shows.

Reference collections agent

Script-first and listen-first versions · written by MPhora

7 vs 3Cut-ins: script-first vs listen-first, six calls

Same virtual borrowers, two agent versions. Listen to each call and see where the conversation turns.

Observed answer

Asking changes what comes out

Asked what happened, one borrower explains a hospital stay. The script-first version only hears that there is a problem.

Study scope

Cut-ins add error

With the same words, pooled DER on the Hinglish calls rises from 10.5% to 13.1% once cut-ins are added, mostly missed speech, on an open-source baseline.

Study scope

Six calls, not a statistic

This shows the method. Measuring an effect means running many calls.

METHOD & SCOPE

How this study was run.

Each borrower is a saved virtual person written in plain words. The same person took a call from both agent versions, which MPhora wrote.

The calls ran as text. A judge model then marked where each borrower would cut in, with a written reason.

Overlap lengths are drawn from human labels in the public Indic DiarBench benchmark. Every line was spoken by an AI voice and placed on one track, so who-spoke-when labels are exact.

Diarization error rate (DER) was measured with an open-source baseline, with and without cut-ins, on the same words.

LIMITATIONS

What this does not establish.

  • Six calls show the instrument, not a finding about any model, company or population.
  • Nobody interrupted live: timing was added after each text call.
  • All call voices are AI-generated, and native listeners have not checked the Hindi voices.
  • The lender, borrowers and agent are invented; this is not a test of any company's product.

FOLLOW THE SOURCE

Study material.

Explore the available sources behind these findings.

Read full study Interactive demo with audio Public research source

YOUR CUSTOMERS. THEIR CONTEXT.

What would they experience?

Explore how your AI handles people’s goals, situations, and unfolding conversations with VIVID.

PULSE / GET IN TOUCH