The original request sounded like a dashboard project. The real problem started earlier.

Every month and quarter, files arrived from different parts of the operation. People had to check labels, periods, score formats and totals before management could compare performance. The reporting worked, but it depended on the same careful handling each cycle.

Audit summary from the completed reporting run. The totals and output counts are verified; the annual handling range is an estimate, not measured savings.

Keep the input people already know

Replacing Excel would have created a second problem. The teams already knew it, and the source files already existed. I kept that familiar input and rebuilt what happened after the files arrived.

First, I standardized the differences that could quietly change a result. Labels that looked similar did not always mean the same thing. Periods and scoring formats also varied. I converted those inputs into one consistent structure before calculating or presenting anything.

Make missing and replaced files visible

Then I added checks before the reports were produced. A replaced monthly upload had to replace the earlier version rather than create a duplicate. Missing periods had to be visible. Every total needed a route back to its source.

The final quarterly scope covered 101 locations across three months. I checked all 303 monthly location totals against the source. The finished package contained 31 PDF reports spanning 948 pages.

The useful bugs looked believable

One quarterly award calculation returned 0.11 because monthly awards had been averaged when the business rule required a different calculation. Another run completed successfully even though the expected second-quarter records were missing. A green completion message did not mean the business result was complete.

Both failures became permanent checks. The award rule received its own test. The system also learned to distinguish “the process finished” from “the expected records are present.”

AI can investigate. It should not own the totals.

AI can help investigate what changed, surface unusual patterns and support better questions. Scores, totals and reconciliation still follow fixed rules, and the source data remains in control.

The recurring handling opportunity is estimated at 44 to 112 hours a year. That range uses 303 monthly records and 31 report outputs, with an assumed two to five minutes of handling per item. It is a workload estimate, not measured time saved or a return-on-investment claim.

The larger lesson is simple: finance automation is not the removal of spreadsheets. It is the removal of repeated uncertainty after the spreadsheets arrive.