Blog
Thinking on safe agentic AI adoption, governance, and the zone model.
The Secret Sauce for Agentic AI Is the Middleware, Not the Model
The biggest reliability gain in agentic AI will not come from trusting models more. It will come from putting deterministic middleware between agents and expensive systems, so autonomy is bounded by hard budgets, whitelisted actions, and audit logs.
Read article →Running Agentic Tools Without Permission Prompts Can Be Safer
Claude Code and Codex both include permission systems that ask before sensitive actions. That feels safe, but it can turn the human operator into the weakest control point. A better pattern is to run autonomous agents inside disposable, tightly scoped environments.
Read article →Alternatives to Codex Computer Use: Safer AI Workflows on a Linux Server
Codex Computer Use brings desktop control to the Codex app, but it is not the only way to automate UI-heavy work. For many teams, Linux server workflows with APIs, Playwright, queues, and human review gates are safer, more repeatable, and easier to govern.
Read article →Don't Let ChatGPT Be Your Bookkeeper Just Yet
A recent Microsoft Research study found that frontier models can silently corrupt documents when asked to edit them directly. That does not mean AI has no place in bookkeeping. It means the architecture matters.
Read article →The Anatomy of an Agent Skill: Markdown, Scripts, and Repeatable AI Workflows
Agent skills are becoming a shared format for packaging repeatable AI workflows. Here is how OpenAI Codex and Claude Code discover them, load them, and use their Markdown, scripts, references, and metadata.
Read article →Case Study: Inbound Lead Qualification and Answer Drafting
How we built a draft-only AI workflow that prequalifies inbound leads, matches each inquiry to relevant services and client context, and creates a reviewed-before-sending Gmail draft, with a feedback loop that improves tone over time.
Read article →Case Study: Accounts Receivable and Payable Intelligence for an Event Management Firm
How we built a read-only invoice intelligence layer for a European event management company, automatically matching hundreds of invoices against bank transactions each month and surfacing only the exceptions that need human attention.
Read article →Case Study: Surfacing Commodity Intelligence for an Investment Firm
How we built a structured AI workflow that synthesises EIA and OPEC reports against an active energy portfolio, turning weekly manual research into a 20-minute automated intelligence cycle with human triage at the centre.
Read article →Designing AI-to-Human Handoffs: Where Structure Creates Value
The interface between the AI Zone and the Human Zone is where the most value is created, and where most AI implementations go wrong. Here is what makes a structured handoff succeed.
Read article →Agentic AI for Family Offices: Managing Risk Without Slowing Down
Family offices operate under unique constraints: lean teams, complex portfolios, and fiduciary obligations that leave no room for AI-related errors. Here is how to build an AI governance framework that works for your structure.
Read article →Agentic AI Zones Explained: Why Separation Is the Safest Path Forward
The biggest risk in AI adoption is not the technology itself. It is blurring the lines between what humans and AI should each be responsible for. Here is how the two-zone model eliminates that risk by design.
Read article →