The spec lived in a Slack thread.
Your team clarified which tables and filters to use. The agent never received that context.
Before the automated ARR report runs every Tuesday morning, there is a Slack thread where the actual spec was worked out. Exclude the test tenants: there's a flag, is_test = true. Exclude the migrated_legacy accounts: they show MRR but they're on a grandfathered plan that Finance doesn't count. Use mrr_contracted, not mrr_monthly: they diverge for annual accounts. The enterprise_segment join needs to go through accounts_segments, not directly to users.
The bot sent the report at 9am. The board deck used it. The context that made the report correct lived in four messages in a Slack thread from the day before.
This is how operational specifications actually work. They emerge from questions asked and answered, edge cases identified and resolved, the back-and-forth between whoever knows the data and whoever needs the answer.
The automated system received the query, not the clarification. When the same query runs without a human in the loop, the clarification isn't there. The system makes its own choices about which accounts to include, which table to use, which join path to follow.
The specification that makes a query correct often lives in the conversation that happened before the query was written.
