AI/LLM Strategy

The AI question every team is wrestling with right now is not “can we use a language model” but “where should we, where shouldn’t we, and what does it cost us to be wrong either way.” The demos are easy. The productionised, evaluated, guardrailed, on-budget feature that a real customer pays for — that’s the work. AI/LLM Strategy is the engagement we built for the moment before a team starts that work, when the question of where to aim is still open.

We start with the product as it exists. We map the surfaces where a language model could plausibly earn its keep — search, summarisation, authoring help, support automation, structured extraction, agentic flows — and we stress-test each one against three questions: what’s the quality bar this surface actually needs, what’s the failure mode when the model gets it wrong, and what’s the unit economics at the volume you expect. Most of the surfaces that look exciting in a demo fail at least one of those questions. We say so, in writing.

Where the surfaces do pass, we go deeper. Model choice — Claude, GPT, open-weights, fine-tuned specialist — and provider strategy, because a multi-provider abstraction on day one is worth ten times what it costs to retrofit later. Retrieval architecture if the feature needs it. An eval harness plan so “it got better” is a measurement, not a vibe. Guardrail design for the real failure modes, not the ones a safety team guessed at. And a rough cost model at the traffic you’re actually planning for, not the traffic the pricing page assumes.

What you get back is a written strategy document — opinionated, specific to your product, with prioritised surfaces and a defensible reason for everything we recommend against. If the honest answer is “you don’t have an AI product here yet, you have two AI features,” we’ll say that. The studio’s brand is intellectual honesty over marketing polish, and AI is the one domain right now where that honesty is hardest to get anywhere else.