Five reasonable choices. Each one wrong.

The agent was asked: how many active customers do we have? It made five decisions. Each one was reasonable. Each one was wrong, for a reason that wasn't in the schema.

Step one: table selection. The agent picked accounts, with 2.1M rows and a clear name. The right table was active_subscriptions, with 41K. Nothing in the catalog flagged the difference.

Step two: join logic. It joined accounts directly to subscriptions. The canonical path goes through accounts_segments. The direct join double-counts. The query ran without error.

Step three: filter applied. It filtered on status = 'active'. Trials, grace-period accounts, and internal test accounts all match that filter. Finance excludes all three. The schema doesn't say so.

Step four: metric definition. It used any login in the last 30 days as the activity signal. Finance requires paid and not in grace period. Twelve thousand accounts fall in the gap.

Step five: result returned. One number. No errors. No caveats. Four undocumented choices determined whether it was right.