# Writing

Dated method posts. Every number in them names who measured it, and each post lists its sources.

- [Attackers now run AI agents. Their target is your keys.](https://www.grayhavenindustries.com/blog/attackers-run-agents) (22 October 2026): In October three hosting platforms said attackers may have reached the keys their customers stored with them, many of them keys to AI services. The incident reports, threat research and agency guidance agree on what to do first.
- [Prior authorization has a clock now. The research is still yours.](https://www.grayhavenindustries.com/blog/prior-authorization-clock) (30 September 2026): From 2026 a federal rule gives many insurers seven days to answer and requires a reason for every denial. The hours go into preparing the request, and that is the part an agent can do.
- [What an AI agent is. And what it may do alone.](https://www.grayhavenindustries.com/blog/what-an-agent-may-do) (30 September 2026): The companies that build the models define an agent as a model that chooses its own steps. The rules for clinics, banks and law firms already say where a person has to stand.
- [Start with the smallest model. Move up only when it fails.](https://www.grayhavenindustries.com/blog/smallest-model-that-works) (30 September 2026): By Stanford’s count, the price of a fixed level of AI performance fell more than 280-fold in two years. For sorting work, a small model trained on your own cases can beat a large general one.
- [Most AI projects stall. The studies say why.](https://www.grayhavenindustries.com/blog/why-ai-projects-stall) (30 September 2026): S&P Global surveyed 1,006 companies and RAND interviewed 65 engineers about projects that failed. The causes they found are older than the models, and a trial can test for each one.
- [Where Jev fits in a regulated system. And where it doesn’t.](https://www.grayhavenindustries.com/blog/decision-layer) (23 September 2026): A new kind of model answers yes, no, or a pick from your list. Here is where it belongs in a clinic, a bank or a law firm, and where it doesn’t, from its maker’s own documentation.
- [Jev says it is 95% sure. Check that on your own files.](https://www.grayhavenindustries.com/blog/ninety-five-percent-sure) (23 September 2026): A new kind of model answers with confidence scores. For a bank, a clinic or a law firm, the question is whether those scores hold on its own files.
- [Seventeen minutes of work. Twenty-two days in the building.](https://www.grayhavenindustries.com/blog/seventeen-minutes) (8 September 2026): An insurer told Michael Hammer that number in 1990. It is still the number that decides whether AI does anything for a back office.

---
Grayhaven Industries · https://www.grayhavenindustries.com/blog