1The situation
What is happening in this world?
The lender (Navya Finance), the borrowers and the agent are invented. MPhora wrote the agent from public descriptions of collections assistants; it is not any company’s product. The calls ran as text on VIVID, then were voiced. Nobody interrupted live: timing was added after the call, with overlap lengths drawn from human labels in the public Indic DiarBench benchmark (Interspeech 2026, CC BY 4.0).
2Who talks to whom
A virtual borrower you can change, and the agent under test
3The call
4What we measure
How the agent behaved, how the borrower felt, and whether a system could hear it
Click a row to open that call. Both views are the mean of clarity, helpfulness, empathy and relevance. Read DER across a row (same call, same words). Do not compare DER between agents: the calls differ in length and content. Six calls show the instrument, not a finding about any model. DER (diarization error rate) is the share of speech missed or given to the wrong speaker, from an open-source baseline. Cut-ins are marked after the call by a judge model.
5How we double-check
Two views of the same call, then people on what matters
All six calls: how closely the two views agree
6Change one sentence
The borrower is a setting, not a recording
We replaced only the last sentence of Arjun's description and ran the same script-first agent. One real call: single calls vary, so measuring an effect means running many. He pushed back and offered ₹3,000, but never asked for a manager: the description shapes him, it doesn't script him.