Your bottleneck was never the query
Analysts spend far less time computing numbers than explaining them. So the temptation is to point AI at the whole pipeline: write the query, read the output, tell the story.
Split that, because the risk profile at each step is wildly different.
Where it's genuinely strong
- Explaining code you inherited. "What does this 200-line query actually do?" is probably the highest-return prompt an analyst has.
- Writing the boilerplate. Window functions, date-spine joins, the regex you look up every single time.
- Turning your findings into prose. You've done the analysis; drafting the exec summary is the tedious part, and it's a pure language task.
- Interrogating your own conclusion. "Here's my read. What would a sceptical CFO ask?"
- Naming what to check. "EU revenue dropped 12% last month. What are the plausible causes — including boring ones like a pipeline change?"
Where it will burn you
- Believing a number it produced without checking it.
- Letting it infer causation — which it will do fluently, constantly, and for free.
- Handing it raw customer data (that's two topics from now).
The rule that holds:
AI is excellent at the words around the number, and unreliable at the number.
Which is convenient, because the words were the slow part.