AI pilots rarely fail because the model was bad

The AI pilots I've seen fail almost never fail because the model was bad. They fail because the data and process underneath it were never actually digitised: three versions of a spreadsheet, a process nobody documented, a team still working around the "official" system.

Before any AI initiative, the honest question is: which of our core operations are still running on spreadsheets and tribal knowledge?

Answering that and fixing it first is less exciting than announcing an AI pilot. It's also the difference between AI that compounds in value and a demo that gets shelved after one underwhelming result.

*First posted on [LinkedIn](https://www.linkedin.com/feed/update/urn:li:activity:7506880413529792514/).*