What analysts actually do with models all day

Capable models are roughly a speedup on the part of analytics work that was never the hard part: writing the SQL once the question is clear.

What did not get faster is turning “how many customers are actually active?” into a question with one answer. That still takes meetings. No model shortens the politics — though one thing helps: putting the ambiguity list in front of the person who asked.

Four definitions of “active”#

Logged in this week. Did a paid action. Opened the app. Has a seat. Pick wrong and the dashboard is fiction with nice charts.

Models write SQL; people write meaning#

Ask a vague question and the model invents a definition. That feels helpful. It is a silent policy change. Force the definition into the prompt or the ticket before any query runs.

Send the ambiguity upstairs#

A good analyst prompt does not hide uncertainty. It returns the list of terms that still need a human decision. That list is the real deliverable on day one.

Sanity checks in the PR#

Row counts, null rates, and “does this match last month’s signed-off number” belong next to the SQL. If the model cannot show its work, do not merge it.

When the schema cannot answer#

Sometimes the honest output is: we do not store that. Shipping a wrong metric because the model filled the gap is worse than saying no.