How we run an AI discovery week.

Every engagement at starmo starts the same way: one week, fixed-fee, before any building. Clients arrive with a list of things they want AI to do. Our job in that week is to figure out which of those things are worth building, which are worth doing differently, and which aren't worth doing at all.

We do four things, in roughly this order.

Listen

We sit with the people whose work would change. Not the executive sponsor, not the head of engineering: the analyst, the support agent, the operations manager. The person who knows how the work gets done, including the fifty exceptions that aren't in the SOP.

AI changes processes, not slides. If we don't understand the process, we'll automate the slide.

Map the workflow

We document one or two end-to-end workflows in detail. What's the input, what's the output, where do humans intervene, where does it go wrong, where is the time spent.

Most "AI opportunities" turn out to live in the boring middle: a routing decision, a data normalization step, an exception that fires once a week. The interesting question is rarely where can we add an LLM. It's where is the bottleneck, and is an LLM the cheapest way through it.

Evaluate

For each candidate, we sketch the simplest version that could ship, what we'd need to evaluate it, and what could go wrong in production. If we can't write down a way to measure whether the AI is doing its job, we don't build it.

Most candidates fail this step. That's good: better to catch them now than at month three.

Recommend

At the end of the week we hand back a short document. Two or three things worth building, in order. For each: a sketch of the system, an estimate, an evaluation plan, and a list of risks. Often a longer list of things we don't recommend, with reasons. Sometimes the recommendation is don't hire us, and we mean it.

The diagnostic week costs less than two days of building. It is the cheapest way we know to avoid building the wrong thing.