Bevaya Says InsurGPT Beats General AI Models on Loss Runs

Bevaya said in a press release that its InsurGPT model reached 93.1% field accuracy across 346 insurance loss runs. The strongest general purpose model scored 85.3%, while all general purpose models scored between 78% and 85%.

InsurGPT scored 86.6% on claim numbers, policy numbers and claimant IDs, compared with 72.8% for the general purpose models. Bevaya also reported scores of 82% for claim type and 89% for claim status, while competing models scored as low as 59% on these carrier specific fields.

The company said InsurGPT was trained on more than 300 million nonpublic insurance documents. Its production system uses a second model to check answers against source documents, assigns confidence scores to individual fields and sends uncertain results to human reviewers. Bevaya reported production accuracy above 98%, with more than 90% of loss runs completed without human review.