Report
A structured AI workflow for empirical analyses
In our own analytics practice, we have built a three-stage AI workflow with approval gates, which makes every analysis verifiable step by step, and have tested its effectiveness in a controlled comparison.
Let’s discuss your project
Contact us directly and let's talk about a possible collaboration.
Background
AI models deliver an analysis in a matter of minutes, but without any evidence of how it was arrived at. Those responsible for studies and expert reports cannot work this way: assumptions remain hidden, results cannot be reproduced, and errors are only spotted at a late stage. We wanted to harness the speed without sacrificing verifiability.
Our Approach
-
1
Breakdown of the analysis process into three stages (analysis brief, data processing stage, evaluation) with approval by the responsible economist at each transition point.
-
2
Carrying out all calculation steps using stored scripts, so that every figure can be reproduced from the raw data.
-
3
A controlled comparison of several working methods using identical test cases, in order to measure the contribution of the structure rather than merely asserting it.
Result
This workflow is used in our projects. Analyses can be traced through their intermediate outputs, assumptions are documented, and results can be reproduced from the raw data. The structure requires additional computational runs, and we disclose this overhead rather than concealing it. We are currently compiling the measurement results into a methodological paper.