Insights
Useful principles and clear opinions on review paths, data boundaries, on-premise deployment, ISO 42001, and what it takes to move AI from pilot to production.
2026-06-27
Most AI programmes stall in the pilot. The gap to production is not a better model — it is review discipline, data boundaries, and a control plane the institution actually owns.
2026-07-10
On-premise AI deployment is not nostalgia for server rooms. For books that carry owner wealth and related-party flows, the data boundary is the decision — and it decides everything else.
2026-07-10
Language models are good with words and unreliable with arithmetic. The fix is architectural, not better prompting: deterministic code computes every figure, and the model only narrates.
2026-07-10
The useful shape of AI in the pre-close window is not auto-posting or auto-blocking. It is a ranked, drill-through exception queue that a human dispositions — read-only by design.
2026-07-10
A family office ledger is not ordinary corporate data — it is the owner's private affairs in double-entry form. AI can help, but only inside a boundary the family controls.