The biggest takeaway here isn't about BigQuery or sales forecasting it's about how to work with AI on real systems.
A pattern we've started following is "sample first, build second." Before asking an LLM to generate SQL, integrations, or transformations, pull a handful of real rows and let the model reason from the actual schema instead of assumptions. It avoids hours of debugging code that's technically correct but built on the wrong mental model.
The other point that resonated is that AI is excellent at accelerating analysis, but domain expertise is still what separates an interesting output from a useful business decision. Models can surface anomalies; people provide the context that turns those anomalies into action.
The biggest takeaway here isn't about BigQuery or sales forecasting it's about how to work with AI on real systems.
A pattern we've started following is "sample first, build second." Before asking an LLM to generate SQL, integrations, or transformations, pull a handful of real rows and let the model reason from the actual schema instead of assumptions. It avoids hours of debugging code that's technically correct but built on the wrong mental model.
The other point that resonated is that AI is excellent at accelerating analysis, but domain expertise is still what separates an interesting output from a useful business decision. Models can surface anomalies; people provide the context that turns those anomalies into action.