01
The initial request was only the first hypothesis.
The business already had Balanced Scorecard and Profit & Loss reporting. A Balanced Scorecard compares performance across several measures; a P&L shows the financial result. The pain lived in everything people had to do after the source files arrived.
Before choosing the software, I reconstructed how the files moved, which comparisons management needed, which numbers were authoritative, where judgment was required and which steps simply repeated.
02
Keep Excel where it helps. Remove the repetition after it.
The upstream teams already knew Excel and the source files already existed. Replacing that part would have created adoption work without equivalent value.
I kept the familiar input and built validation, standardized reporting data, performance views, role-aware access and repeatable exports around it. Management could move between locations, periods and financial views without reconstructing the analysis each cycle.
03
A correct-looking screen is not enough.
The final quarterly scope covered 101 locations across three months. All 303 monthly location totals matched the authoritative source. The reporting package contained 31 PDF outputs spanning 948 pages.
Testing also exposed real mistakes in quarterly scoring and upload replacement. Fixing the rules and rerunning the data audit was part of the work. The verification mattered more than whether the dashboard looked convincing.
04
The files did not match, so I fixed that before calculating anything.
The source workbooks were designed for people, not a reporting pipeline. Labels, periods and score formats could look similar while meaning different things. I translated them into one explicit data shape before calculating or presenting anything.
That layer made the system easier to test and gave every report a traceable route back to the original input. It also meant a replaced upload could be processed as a replacement, rather than quietly creating a second version of the same month.
05
The useful bugs became permanent tests.
One quarterly award calculation produced 0.11 because monthly awards had been averaged when the business rule required a different aggregation. It was plausible enough to pass a visual glance and wrong enough to change the result.
Another production run completed successfully while the expected second-quarter records were absent. Both failures changed the system: the scoring rule became an explicit regression test, and process completion was separated from proof that the expected business records actually arrived.
06
Use AI for investigation, not arithmetic authority.
Scoring, calculation and reconciliation follow fixed rules. The source data remains in control. AI sits above that structure to help investigate patterns and ask better questions about what changed.
That separation keeps a useful assistant from becoming the calculator, the database or the person who decides which number wins when systems disagree.
07
Model the opportunity without pretending it already happened.
The recurring workload model uses 303 monthly records plus 31 report outputs, at an assumed two to five minutes of handling per unit. That produces roughly 11 to 28 hours per quarter, or 44 to 112 hours per year.
This is a transparent estimate of the handling opportunity. It is not a stopwatch study, a measured saving or a financial return claim.