Insights
Briefings and guides for the people deciding how AI agents will operate inside a regulated estate — what sovereignty means in practice, where models belong, and what evidence has to exist afterwards.
Guide
Agents now pull data from email, CRMs and document stores and send it to models at machine speed. Where that data goes, what can go wrong, and why retention, controls and audit are now the whole game for regulated organisations.
Read the briefingAnthropic's own filing concedes that open-weight models can take enterprise work where cost beats capability, and that the durable advantage may be compute rather than weights. Both readings have consequences for anyone buying AI.
2 November 2026 · 5 min read
ReadZDR is a useful policy promise, but it is still a promise about data that already left your environment. The strongest privacy control is not sending it in the first place.
14 September 2026 · 15 min read
ReadData residency is the easiest part of sovereignty and the least useful on its own. The harder questions are about infrastructure, models, policy and evidence.
14 July 2026 · 16 min read
ReadFrontier models are a reasonable choice for many tasks. They are a poor foundation for the internal processes that define how your organisation works.
30 June 2026 · 15 min read
ReadAgents need the same treatment as any other privileged workload: identity, scoped authorisation, mediated tool access, and a record of what happened.
12 June 2026 · 17 min read
ReadA practical routing model: self-hosted open models for internal processing and sensitive content, commercial frontier models where capability justifies the exposure.
28 May 2026 · 16 min read
ReadAssurance rarely turns on architecture diagrams. It turns on whether you can reconstruct specific decisions, months later, without asking a supplier for help.
9 May 2026 · 15 min read
ReadMost agent pilots succeed and then stall. The reasons are consistent, and nearly all of them are decided before the pilot begins.
22 April 2026 · 16 min read
Read
We'll run a governed pilot on a real workflow in your environment and hand you a board-ready summary of risk, cost and results. A human replies within one working day.