Insights topic
AI strategy & economics
The expensive decisions are rarely technical. These pieces cover strategy framing, delegation depth, and the cost assumptions nobody wrote down.
AI strategy usually starts by answering the wrong question
Boards ask which AI strategy to adopt when the more useful question is which decisions in the business are worth delegating, and to what degree. Framed as a technology choice, AI strategy collapses into a model or vendor bake-off; framed as a delegation problem, it forces the harder conversation about which judgment calls the organization is willing to hand to a system with a known error rate, and which ones it isn't, regardless of how good the model gets. That distinction is what separates AI programs that survive contact with a bad quarter from ones that get quietly shelved the first time an agent makes an expensive mistake in public.
Autonomy depth is an economic decision as much as a technical one
How much autonomy to grant an agent is a delegation problem with real economics attached: more autonomy generally means fewer humans in the loop, faster throughput, and a higher blast radius when something goes wrong. The cheapest model that's right enough for a task is very often the correct strategic choice over the flagship model, because inference cost compounds at volume in ways a pilot never reveals — the unit economics that looked fine at a hundred calls a day look different at a hundred thousand. Strategy work here means matching model capability, autonomy level, and cost per decision to the actual value of that decision, not defaulting to the most capable option available.
You underwrote today's inference price — that's a live assumption
Every AI business case implicitly underwrites today's inference pricing and today's model performance holding steady, and neither is a safe assumption. Prices have fallen before and could rise for specific capabilities as compute gets rationed; a workflow economical today at current API rates can become marginal if pricing shifts, and the reverse can also unlock use cases that weren't viable last year. Durable AI strategy treats cost per decision as a variable to be monitored and re-underwritten, not a constant set once at launch, and builds in the flexibility to swap models or scale autonomy down if the economics move against the business case.
Why “AI strategy” is usually the wrong question
Most organizations asking for an AI strategy don’t need one. They need a way of deciding — a small set of questions asked per decision, not a plan written once and outdated by spring. This is the frame the rest of our thinking expands from.
Read
Autonomy is a delegation problem
The agent revolution is real: systems are moving from advising to acting. But the framing — autonomy as doing more with less human friction — hides what actually changes. The moment a system acts instead of advising, it is making decisions. You haven’t bought a faster worker; you’ve delegated authority — and everything hard about agents lives there.
Read
AI model selection is a habit, not a decision
Reaching for the largest available model is a sound answer to a question you cannot yet answer any other way, and we would rarely talk a client out of it at the start. But the default outlives the uncertainty that justified it. What began as insurance quietly becomes the reason nobody ever measured how well the system works.
Read
The assumption nobody wrote down
Building the case on falling AI inference cost was the empirically supported call, and we would rarely have argued against it. What concerns us isn't the assumption. It's that it was never written down as one — so nobody has worked out what breaks if it stops holding, and nobody owns it.
Read