AI strategy · fractional leadership · speaking

Advisory

Strategic and fractional AI advisory for software companies and their investors — for the projects where the binding constraints turn out to be regulatory, architectural or economic rather than technical.

Three engagement shapes, all deliberately narrow. Strategic advisory for leadership teams and investors deciding what to build, what to buy, and what a generative feature will actually cost in production — short engagements that end in a decision along with a roadmap. Fractional leadership for companies that need senior technical judgment without a full-time hire. Speaking for audiences who would rather hear from someone still doing the work.

The question is rarely which model is best. It is which one is good enough for the job in front of you — and that answer decides where the work can run: on-device, in your own tenancy, or on someone else's API. What a finished job costs, what the model is ever allowed to see, and whether the design survives a compliance review all follow from it.

And the bar keeps getting cheaper to clear. Work that needed a frontier API last year clears it on open weights today, so the design worth having is the one that lets you swap the model underneath without rebuilding around it.

In practice

Is this generative feature economically viable at your price point — before it is built?
Which model is good enough for the job, and what it takes to swap it when a cheaper one clears the bar.
Where the model runs — on-device, in your own tenancy, or on a third party's API — and what each choice costs in money, latency and exposure.
Architectures where sensitive data never reaches a third-party model — designed in, not bolted on afterwards.
Build-vs-buy calls, and diligence support: a technical read on a software asset, or on the AI claims in a deck.
Fractional senior technical leadership where a full-time hire is the wrong shape.

What this is built on

Not a forecast. The Frontier Rate Card prices every frontier-class model across ten ordinary tasks — what a finished piece of work actually costs, rather than what a token costs. It is published, current, and the same analysis that sits behind the first question above.

Speaking

Talks come from current practice — what shipped, what it cost, and what turned out not to work. The material is rebuilt from the published work rather than revised from a deck, so it dates in months rather than years.

What frontier AI actually costs per finished job — and why the token price is the wrong unit.
Good enough: choosing a model you can afford to replace.
Where the model runs, and what each answer costs you.
Reading the AI claims in a deck.

Talk to us about this →