# Blog & notes

Short notes on what actually works when you put AI into insurance and other places where mistakes cost money. Most link back to a project.

- [Is your company visible to ChatGPT? A practical AI visibility check](https://kivilaid.ee/blog/ai-visibility-check) (23 September 2026): More buyers ask an AI assistant instead of searching. Here is how I check in twenty minutes whether the assistants know you exist.
- [AI audit or discovery sprint: which one do you actually need?](https://kivilaid.ee/blog/ai-audit-vs-discovery-sprint) (2 August 2026): Most companies buy the wrong first AI engagement. The real question is what kind of evidence you need next.
- [Custom AI builder or agency: when production systems beat retainers](https://kivilaid.ee/blog/custom-ai-builder-vs-agency) (2 August 2026): The real question isn't who runs the kickoff workshop. It's who owns accuracy, latency and governance after it.
- [Why underwriting AI fails without provenance](https://kivilaid.ee/blog/underwriting-ai-provenance) (2 August 2026): Underwriters don't need a clever model. They need to see where every extracted value came from.
- [Sub-100 ms pricing is a governance problem](https://kivilaid.ee/blog/sub-100ms-pricing-governance) (2 August 2026): Fast ML-augmented rating only works when actuaries own the bounds.
- [MCP as the integration layer for public data](https://kivilaid.ee/blog/mcp-public-data) (2 August 2026): AI assistants need live facts, not stale training data. One hosted MCP server fixes that for Estonian public data.

---

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