Writing on sovereign AI, deterministic decisioning, and what automating a regulated process actually costs.
31 July 2026 · Cost of intelligence · 9 min
Fifteen chained LLM calls at 90% accuracy each is 20.6% end to end. Tokens turn out to be 36% of the bill and the largest line is the errors nobody caught. A full cost-of-ownership model for document-heavy decisioning, with a calculator you can argue with.
More pieces moving here shortly.
