Scores
One row for each measure, over the normal cases in this experiment.
| Measure | Result | What it means |
|---|---|---|
| Field accuracy | 100.0% | Share of facts pulled from the documents that match the answer key. |
| Appetite agreement | 90.0% | Share of normal cases where the AI and the underwriter chose the same appetite, meaning whether the insurer wants the risk. |
| Accepted as is | 85.0% | Share of normal cases the underwriter could approve without an edit. |
| Corrections per case | 0.15 | Average number of edits an underwriter would make to one normal case. |
| Schedule within 2 points | 95.0% | Share of normal cases where the schedule change is within 2 points of the underwriter's. |
| Premium error, median | 0.5% | Typical gap between the AI's premium and the underwriter's premium. |
| Latency, median and 95th | 13.8 s and 17.3 s | Time from the submission arriving to the draft being finished. |