AI Data Projects Are Hard — But Not Because of the AI
Yes. And the hard part sits in the layers underneath the AI, not in the AI itself.
A client will typically say something like: “I’ve got these data tables — A, B, C, D, E, F — and I want insights out of them on H, I, J, K, L, M, N, O, P, Q.”
It sounds like a simple matching exercise. In practice, it’s three steps.
First, you actually have to get fluent in the A-through-F tables — every field, and how the tables relate to each other: primary keys, foreign keys, join conditions.
Then you have to align with the client on exactly what they mean by each of those H-through-Q metrics: the definition, the calculation formula, the scope, the time granularity.
Finally, you take those definitions back to the A-through-F data and check them field by field — can the data actually support what’s being asked?
Most of the time, you’ll find that a lot of the metrics simply can’t be built, or only a shallow version of them can. The kind of drill-down and multi-dimensional cross-analysis the client is picturing — call it R through Z — is often just not achievable with what’s there. So this is the point where you need to manage expectations early, and sort every metric into one of three buckets: directly supportable, supportable once you add data or redefine the metric, or not supportable at all right now.
This whole stretch is delicate work. It’s a real test of how well the project team coordinates, and of how tightly someone is holding the whole project together. Skip that, and it’s easy to fall into one of two traps: obsessing over the exact definition of a single metric while losing sight of the overall metrics framework, or obsessing over the framework while losing the business meaning underneath it.
Only once you bring AI into this do you add another layer to think about: can the data actually be understood correctly by a model? Everything an analyst used to carry around in their head — what a field really means, how a metric is really defined, which exceptions apply — now has to be written down in black and white, as a data dictionary and a semantic layer.
What this needs is people who think clearly. There just aren’t that many of them.