An outdoor-cooking manufacturer asked its warehouse 'what changed this week' and got an answer with sources

The data team had the pipelines and the weekly business review. What the executives wanted was the why. We built a semantic layer over the warehouse and an agent trained with verified queries on the questions the CEO asks every week: what changed, why sales moved, which products and partners are growing or at risk, which channels perform.

What was held back on purpose

Three capabilities were kept out of the first demo rather than shown half right: budget against plan (a duplication bug, nearly fixed), promotional sensitivity (the company runs near-continuous promotions, so isolating one promo's effect is unsolved) and SKU cannibalisation. Saying what is not ready is part of the method.

How the answers get reliable

AI over data is right about 80 to 85% of the time out of the box. Reporting needs 99. The difference is engineering: a semantic layer with context at the use-case level, 30 to 40 verified question-and-answer pairs per domain, and an evals layer that catches drift. A thumbs-down from a business user turns into a new verified query, so the system improves with use rather than needing maintenance.

The gut check came from the data lead: a product he knew had been a disaster was flagged as a problem without being told. A new product with no budget and no prior-year sales exposed a real modeling question: what is the baseline?