Our proposalData advantage
Our proposal · data advantage

Custom models need data. Policybazaar has plenty.

Policybazaar's data

~10 crorecustomer interactions a month
  • 21 lakh

    calls a day

  • 16 lakh

    minutes of conversation a day

  • 7 lakh

    documents a day

  • 2.7 lakh

    emails a day

  • 17,000

    chats a day

  • 500 GB

    of data a day

Source: PB Fintech Analyst Day, 11 May 2026, slide 52.

PB Flash and PB Pro

PB FlashThe everyday model

Takes most messages: details, quotes, forms, service. Fast and cheap

PB ProThe expert model

Takes the hard few: which plan suits Papa's diabetes, what a clause means.

How we make them
  1. 1Learn from your chats, calls and outcomes
  2. 2Reward the right plan, the right price, a compliant answer
  3. 3Ship only when it beats the frontier on your evals

The evidence

Trained on one workflow, a smaller model beats the frontier.

IllustrativePolicybazaar. Hard health turns: which plan suits a condition, what a clause means
Cost per turn · cheaper →Quality on your evalsFable 5 · 6.6 sOpus 5 · 6.4 sSol · 4.8 sSonnet 5 · 4.1 sGemini 3.6 Flash · 4.7 sGLM 5.2, open · 2.1 sPB Pro: GLM 5.2, post-trained · 2.1 s, fastest
Dashed line: Fable 5. Quality: the shape we'd aim for; real numbers come from evals on your own chats. Seconds: measured 11 Nov, median of 5 runs of one hard pb.ai turn. GLM 5.2 on a speed-optimised host, as Dvar would serve it; the others on their default endpoints
Above Fable 5
The bar PB Pro must clear
About 1/10
Of Fable 5's cost per turn
3.7× faster
3.1 s vs 11.4 s a turn: pb.ai's 31 test chats, open models vs Gemini 3.1 Pro, measured

How: Fine-tune GLM 5.2 on pb.ai's good conversations, then reward the right plan, the right price and a compliant answer