The Coca-Cola Company, Summer 2026
Replacing a manual data pull at Coca-Cola
An internal tool that replaced a four-person, three-day manual pull.
$ run brand-review plan query, schema-checked ok compute 25 to 30M rows ok reconcile Excel, BI, Python ok reconcile one figure HALT > two systems disagree > run stopped, mismatch reported $ _
A recurring analytical request took a four-person team about three days to pull by hand. I interviewed 10 of the people who relied on it, then built an internal tool to replace the manual pull.
Azure OpenAI plans the query. Deterministic Python computes every figure over a 25 to 30 million row backend. Model output is schema validated, and more than 50 reconciliation checks run across Excel, Power BI and Python.
Along the way I traced a recurring discrepancy to the order in which figures were rounded, and wrote the calculation standard the team adopted.
What I found
If two systems disagree on a number, the tool stops and reports it instead of picking one. The checks caught nine mismatches.
| Manual pull it replaced | 4 people, ~3 days |
|---|---|
| Direct users | ~50 |
| Rows in the backend | 25 to 30M |
| Reconciliation checks | 50+ |
| Mismatches caught | 9 |
| Employees reached by its output | 500+ |
The tool, its code and its data are internal to the company, so the window above is an illustration of how a run behaves, not a screenshot.
Next case study: Cutting support tickets at Quadcore