every conversation i'm in this year starts the same way. what's our ai play. the board has it on the agenda. the cfo has a line item for it. someone three levels down has been asked to figure out which tool to buy. and what makes the question hard isn't picking the tool. the tool feels like the question. it usually isn't.
every successful ai project i've seen started where a much harder project had already ended.
the discovery sessions i run at ai compass usually take six or eight hours — long enough to map the actual workflows, not just the org chart. by the time we're an hour in, it's usually clear whether a company is going to ship anything with ai or just spend a year talking about it. it has nothing to do with technical sophistication. it has very little to do with budget. it doesn't even depend on what model they end up using. it depends on whether — before the model arrives — the company has already done the harder work.
the harder work has a hundred names depending on who's describing it. operational maturity. process discipline. organizational clarity. what i call it doesn't matter. what matters is that the work is not technical. it isn't a project anyone can quote you a price for. it's the quiet, unglamorous job of making the company explain itself to itself — so clearly that a stranger could walk in and run a piece of it without asking anyone a question.
- 01 data someone wrote down the answers before the model needed to look them up. definitions, decisions, policies — not just numbers in a warehouse.
- 02 workflow the process exists somewhere other than the head of the person who's been doing it for eleven years.
- 03 ownership when the outcome moves, there's a name attached. when it doesn't, the same name is the one who has to explain why.
- 04 measurement the number was named before the project started. people will be able to point at q3 and say "it worked" or "it didn't."
- 05 adoption one workflow that used to be done a certain way is now done a different way. you can name it. you can date it.
foundations work is harder than model work for one simple reason: model work has the comfort of being technical. there's a vendor to call, a benchmark to chase, a model card to read. foundations work is operational, political, and slow. the people who can do it are usually the same people most needed to keep the lights on. there's no demo for it. there's no slide that summarizes it. there's no panel discussion at your industry conference called how we documented the lead-time tolerance for our top eighteen suppliers.
the companies that skip foundations spend more on ai, not less. they buy more tools because no single one ever fits. they run more pilots because each pilot tries to do the foundational work disguised as model work. they get further from the answer with every quarter that passes. by year three they have a presentation called lessons learned from our ai journey and the lesson is always — written carefully so as not to indict anyone — that they should have started somewhere else.
the companies that do the foundations work first don't necessarily ship more ai. they ship less theater. some realize the model was unnecessary in the first place. some ship one quiet, durable thing that nobody outside the company hears about. but the gap between those companies and the rest widens with every passing year, because the foundations they built are reusable — for ai, for automation, for whatever comes next. the model gets old. the foundations compound.
build the system that wouldn't need ai. then add ai if it still helps.