The Warehouse Can Read All 300 Of Your Measures. It Still Can't Tell Which One The Room Decides On.

The Four Skills, part 1 of 4 | Attention

A series on the four skills that matter most for analysts now that frontier models and open-source tooling do the tool-native half of the job. Issue #16 showed the warehouses ingesting the semantic model and skipping the dialect. These four are what they skipped: attention, decisions, vision, systems.

The file was 4 KB.

That is the detail from the Cube study worth sitting with. Rumiantsau and Fokeev put 99 retail questions to three frontier models, once with the warehouse schema and once with the schema plus a markdown file an analyst wrote by hand. The file was worth 17 to 23 points of accuracy. With it, the three models were indistinguishable. Without it, they were also indistinguishable, twenty points lower.

Four kilobytes. About 700 words. A short email.

The schema it sat beside described every table and every column in the warehouse. The file described a fraction of that, and the fraction did the work. Nobody in the paper says how the analyst chose what went in. They just chose. That choice was the experiment, and the models were the control.

Now look at what happened on August 18. Snowflake’s Semantic View Autopilot takes a Power BI model and writes every measure it can parse into a semantic view. Databricks’ Genie Code has done the same since June. Both port the definitions with equal weight. Gross revenue and the helper measure someone built in 2022 to fix a legend arrive in the warehouse as peers. Three hundred measures in, three hundred measures out, each one as important as the next.

That is exactly how nobody in a forecast call thinks.

Sit in one. The deck has forty slides. The room decides on three numbers, and the CFO looks at one of them first. Everyone in the building who has been in the room more than twice knows which one. It is not in the model. It is not in the documentation. It lives in the person who learned it by watching where the eyes went.

That knowledge has a name. Salience. Which number matters, to whom, on which morning. It never lived in the semantic model, which is why ingestion cannot move it, and why the 4 KB file worked. The analyst who wrote it had been in the room.

Here is the uncomfortable version. For most of a BI career, the way to be indispensable was to know more measures than anyone else. The person who could explain all 300 was the person who got the call. The warehouse can now explain all 300. What it cannot do is tell you the three, and the model that is asked a question by a VP next quarter will not be able to either. It will answer the question it was asked. Whether it was the right question is somebody’s job, and the somebody is not a vendor.

Every ingestion tool released this summer treats definitions as flat. Every decision made in a room treats them as ranked. The gap between those two is the skill, and it is not on any certification.

The action. One, and it fits in a Friday afternoon.

Write your own 4 KB file. Twenty metrics, in prose, no formulas. For each one, two things: the decision it feeds and the person who makes it. “Net new ARR. Feeds the Monday hiring call. Decided by the VP Sales.” That is a complete entry. If you cannot name a decision, do not invent one. Move the metric to a second file. Call the second file “computed.”

You will have a long “computed” file. That is the result, not the failure. It is the list of everything your model calculates that nobody uses to choose anything, and it is the list the warehouse ported this month with the same care it gave the twenty.

The first file is the one worth twenty points.