Concept demo

One page to forward. Print or save as PDF.

×Concept demo

Aditya Birla Capital Servicing Demo

We tested 9 AI models on 300 loan and insurance servicing messages. gpt-oss-120b, an open model run by Dvar in India, got 97.2% fully right, above a 95% bar, at 8.6× lower cost than Gemini 4 Argon.

97.2%fully right: above the 95% bar (Gemini 4 Argon 99.2%)
8.6×lower cost per message than Gemini 4 Argon
100%fraud messages flagged (Gemini 4 Argon 100%)

At 10 lakh messages a month (an assumption): ₹1,39,304 with Gemini 4 Argon vs ₹16,200 with gpt-oss-120b per month.

Today

gpt-oss-120b clears the 95% bar at 8.6× lower cost than Gemini 4 Argon, and is more accurate than Gemini 3.5 Flash Lite.

Next: fine-tune for 2 topics

Still under 95%: Policy change 90.6%, Suspected fraud 94.9%. Training it further on your own servicing messages, then re-testing against the same bar.

What was compared

300 synthetic servicing messages in English and Hinglish, many with more than one request, sent 3 times to each of 9 models: leading paid models from Google, and open models run by Dvar in India. Same instructions for every model, one answer key checked by hand, costs at list price. "Fully right" means every request, the product and any urgent flag match. The 95% bar is ours for this demo; your team would set the real one. No model was tuned.

Next step

  1. A 20-minute walkthrough with your servicing or AI CoE lead.
  2. The same test on around 200 anonymised servicing emails, inside your environment, with your team setting the quality bar.

Live demo: https://abcl-demo.dvarlabs.ai

Concept demo · synthetic data · no Aditya Birla Capital customer data · internal benchmark. Run merged-20261021T061351Z, test split, 3 runs per model.