# Why underwriting AI fails without provenance

_2 August 2026 · 2 min · insurance_

Underwriters don't need a clever model. They need to see where every extracted value came from.

Brokers send messy PDFs. Underwriters need to trust every single field that comes out of them. That gap is where most document AI projects fail.

## What production actually needs

- Coordinates back to the source page for every extracted value, so an underwriter can check it in one click.
- A human in the loop for the edge cases, instead of pretending there are none.
- Accuracy measured against gold labels — not F1 scores from a demo on clean samples.

## Hold the launch

At Insly we didn't launch until we reached 99.4% field match on 800 sampled fields across 4,200 submissions. Waiting for that number was the point: it's what underwriters needed to see before relying on the system.

Related: [See Nora, Insly's AI for underwriting teams](https://kivilaid.ee/projects/nora)

---

Source: https://kivilaid.ee/blog/underwriting-ai-provenance · Contact: https://www.linkedin.com/in/andokivilaid · Overview: https://kivilaid.ee/llms.txt