Blog Post
From a Four-Hour Month-End to Ten Minutes
The most important thing an automation tool does is know when to stop
After the finance report automation went live, the client's accountant said month-end went from four hours to about ten minutes. Nice number. But it's not what we most want to write down.
Fast is easy. Correct is hard.
Merging a few spreadsheets, reconciling and summarising them isn't complicated in pandas. The hard question is what the tool does when something in the data is slightly off.
The dangerous case isn't a crash. It's no error, and a wrong result. One wrong figure in a financial report is far worse than a report that's a day late.
What we did
- Every step can be checked on its own: import, clean, match, summarise and export run separately, each with a preview;
- Stop on anything unfamiliar: a changed file format or a missing column halts the run with a message — no guessing;
- List what doesn't match: unmatched transactions aren't forced into place; they're flagged for a person to review;
- Keep a run log: when it ran, which files, how many rows.
A real stop
A few months in, the bank changed its export format. The tool stopped at step one with "unrecognised file format". The accountant messaged us, and the rules were updated within half an hour.
Had the tool cleverly guessed where the columns were, that month's report would have been quietly wrong.
The point
Automation doesn't mean nobody looks. It hands the repetitive part to the program and leaves the judgement calls to people.