Building AI on bad data is a skyscraper on sand

As somebody who does consulting on data and AI topics (drop me a line) and works in the data field for over two decades, I see this happening more and more often.

Great sales presentations and pilot projects are followed by huge failures in production. Even worse, those projects often run way too long because of sunk cost fallacy. At some point people just stop them silently and pretend it never happened.

The root cause in most cases is the same. Companies try to build AI projects on top of huge piles of unstructured and undocumented data. Cheap storage and the popularity of data lakes over the last 10–15 years led to this: most businesses have collected data under the false assumption "we will generate insights later". They didn't invest into infrastructure, governance (documentation and testing), so I doubt this "later" ever comes.

Starting with AI in this case is literally building a skyscraper on sand.

Fix the basics: define data ownership, improve quality, invest into data engineering, and AI will multiply your capabilities.