A founder comes back from a conference sure the company needs an AI team. HR posts a role for a Head of AI, then five more underneath it once the first hire asks for a team of their own. Six months later there is an org chart: a manager, two data scientists, an MLOps hire, someone titled "AI strategy."
Every department has heard the news and sent a request. Sales wants a forecast. Ops wants a scheduling tool. Finance wants the reconciliations nobody enjoys doing by hand. The team logs each one and starts a little of all five at once. Saying no to a department head feels like the wrong way to begin.
A year in, there is a slide with five logos on it. Nothing is running in production. The team is the busiest group in the building and has shipped less than one good specialist could have shipped alone.
Why we build it this way
"Get serious about AI" tends to read as headcount, because a department shows up on an org chart in a way that three working use cases do not. A board can see five new hires. It cannot see the thing that actually mattered, which was never the number of people. It was whether any one of them could take a business problem all the way through to something running.
A team that tries to answer every request becomes a queue, and a queue has no way to say no. Nobody wants to be the person who tells finance to wait, so everything moves forward together, a little at a time, at the pace of the slowest thing on the list. A year later nothing is finished because nothing was ever the actual priority.
We build the opposite: a small group senior enough to go deep on one problem at a time, spanning architecture, applied ML, knowledge engineering, platform and governance — not five generalists who each do a bit of everything and finish none of it. The first use case is expensive, because the pattern does not exist yet. Depth is what makes the second one cheaper: the retrieval layer, the evaluation set, the approval flow, most of it gets reused instead of rebuilt. Hire the large team before that pattern exists, and the coordination cost arrives with none of the reuse to pay for it.
Headcount is what a department looks like from outside. What a use case needs is depth, and depth is not something five generalists add up to.
How it works, roughly
We start from what is already sitting in the building: the requests already in inboxes and spreadsheets, the systems people already work around by hand, the two or three problems leadership already wanted solved before anyone said the word AI. From that we help stand up:
- A charter naming who the team answers to, and what it may decide on its own without asking
- A role blueprint across architecture, applied ML, knowledge engineering, platform and governance, instead of a stack of generalist titles
- A single intake backlog with a written rule for what gets funded, so a department head can see why their request is waiting
- A component catalogue, so the second use case starts from what the first one already built instead of a blank page
- A habit of handing pieces off to your own engineers as soon as they can hold them, measured by how little the team is needed a year in
When someone disagrees with what got funded, the backlog rule is the argument, not a preference in the room. The sponsor weighs the case against it and decides. If a new priority genuinely outranks what is running, the rule changes on the sponsor's say-so, not on whoever asked loudest.
A year in, the company does not have a department. It has three or four things running, a written rule for what gets built next, and a small team that keeps shipping because it was never coordinating five requests at once. The call on what comes after stays with whoever holds the charter inside the company — a person decides, and the record of why stays in their own building.
Read about AI Transformation →