AI strategy

The hardest part of an AI project is rarely the model. It is deciding what to point it at, agreeing what success would look like, and being honest about whether the result beats the process it replaces. Teams that skip those steps ship impressive demos and quietly retire them two quarters later.

These posts are about that earlier work. Choosing a first use case with a measurable outcome. Estimating cost properly, including the parts that are not per-token — evaluation, review, support, and the engineering time to keep it running. Deciding between an API, an open-weights model you host, and not using a model at all. Structuring a pilot so that a negative result is still worth having.

There is also a recurring theme about sequencing. Most organisations get more out of fixing their data access and internal tooling than from an ambitious model project, because the second is bottlenecked by the first. Knowing which of those you are actually facing saves a great deal of money.

Written for people who have to justify a decision to someone else — a founder, a manager, a client — rather than for people who only have to satisfy themselves that something is technically interesting.