The problem
Analysts spent most of their time writing the narrative around numbers they had already computed. Drafting was slow; the analysis itself was the fast part.
AI & analytics
A controlled workflow for turning a business question into a drafted analysis — with the validation step that decides whether the draft is usable, kept firmly in human hands.
Portfolio demonstration — illustrative figures, real structure
Analysts spent most of their time writing the narrative around numbers they had already computed. Drafting was slow; the analysis itself was the fast part.
The analytics team drafting commentary, and the reviewers who sign off before anything reaches a business reader.
What the commentary in a report says — but only after a human has checked every claim against the underlying figures.
The validation step is not optional and not automated.
The model never reads raw data. It is given the same governed measures the dashboards use, which is what makes a claim checkable at all.
A sample review cycle. The failure modes matter more than the pass rate.
| Check | Result | Action |
|---|---|---|
| Figures match source measures | 100% | Passed |
| Direction of change stated correctly | 98% | Passed |
| Causal claims supported by data | 82% | Edited by reviewer |
| Material context omitted | 2 cases | Returned to draft |
| Recommendation invented | 0 cases | None found |
Causal claims are the weak spot: a drafted narrative will happily explain why a number moved when the data only shows that it moved. That row is why review is mandatory.
The parts that carry over to a production deployment.
The draft is generated from governed measures and their definitions, not from raw tables, so every figure in the text maps to a measure that can be recomputed.
Each numeric claim carries the measure and period it came from, which is what makes a line-by-line check fast enough to actually do.
No draft publishes without a named reviewer. The workflow records who signed off and what they changed.
The workflow drafts description and comparison. Recommendations and causal explanation are written by the analyst, not generated.