Where running AI meets the bill.
Our point of view on operating agentic AI in production — written by the people who run it once the demo is over. Capability is getting cheaper by the quarter; what stays scarce is the discipline to run it well once it’s live. These are our arguments, our analysis, and our open questions — no vendor’s pitch, and grounded in what it costs to run.
Capability is commoditizing. Accountability is the moat.
Every model gets stronger and every vendor claims to stay. When the tool itself is no longer the differentiator, the scarce thing is someone who will put their name on the result — measured, owned, and answerable when it slips. We argue the durable advantage is no longer what an AI can do, but who is accountable for what it actually does in production.
Read the argument →Why most AI pilots die after the demo
The pilot dazzles, the deck circulates, and then nothing ships. The failure is rarely the model — it is the absence of anyone owning the unglamorous work of running it once the applause fades.
Read → EconomicsAI unit economics, and the bill nobody modeled
Token spend behaves nothing like a software license. We make the case that treating inference cost as a first-class operating metric — not an afterthought — is what separates a margin from a surprise.
Read → OperationsOperating agentic AI is a different job than building it
Shipping a system and standing behind it for the next two years draw on different muscles. We argue that the operating discipline — observability, escalation, drift — is the part the market keeps underpricing.
Read → GovernanceAssurance a board will actually accept
A model card is not an audit. We look at what it takes to give a board defensible confidence in an autonomous system — the controls, the evidence, and the line of accountability behind every decision.
Read → StrategyThe case for the accountable operator
When the people who build your AI also sell the platform it runs on, whose outcome are they optimizing? We argue that independence is a structural feature, not a nicety.
Read → GovernanceProgressive autonomy: earning the right to remove the human
Full automation on day one is a liability dressed as ambition. Our view is that autonomy should be earned plane by plane, with a human floor that never disappears entirely — and we explain why that floor is a design choice, not a weakness.
Read →Begin with a Charter.
A fixed-fee diagnostic that turns these arguments into a plan for your operation — scoped, costed, and run by the people who would operate it.