There is one thing that almost never appears in an AI case study: a failure.
I kept researching during my book project and afterwards. I had conversations about cancelled projects or implementations that cost more than planned. I did not really find what I was looking for.
There is no shortage of case studies across industries and applications, spread throughout the various guides. And they describe exclusively successes. The improvements run at 25, 35, 40, 60 percent. Implementation takes three to six months. Everyone is excited and the numbers speak for themselves. Is that reality?
In real practice, projects fail. Not all of them, but some. Serious management consultancies put the failure rate for AI implementations at over 50 percent. Many claim behind the scenes that most AI projects never leave the pilot phase. The consulting firms whose clients read these guides know that.
But in the published examples this reality does not exist. All you find is the success story where everything works out perfectly.
What interests me is not the question of whether the people writing these reports are lying, because I cannot imagine that they are lying. I think it is more likely that they are reproducing a deeply ingrained pattern that is standard in the tech consulting industry.
In reality it works like this: a consulting firm supports an AI implementation and if it works, it becomes a case study. If it does not, nothing comes of it and the failure disappears. Not actively suppressed but quietly swept under the rug.
The result is an industry that only knows success stories. Not because there are only successes but because failures are not a viable business story. No consultant wins a contract by telling the story of the last client who failed. So they select. And the selection creates a picture that is not necessarily wrong but incomplete.
But whoever only sees successes misjudges the risk and plans with numbers that are not representative. Whoever then fails has no reference material to understand what went wrong because it is documented nowhere.
The anonymity of the examples only makes it worse. When no name is given, there is nothing to check. And nobody asks anyway. So no one will ever learn whether the numbers still hold two years later or whether the costs paid off. Anonymity protects no one and helps no one. It just hides, and that is convenient.
I know consultants who tell me privately that most AI projects fall significantly short of expectations. That data quality was not sufficient, resistance in teams was underestimated and costs exploded after the pilot is what gets told over beers in the evening.
What does that tell us about an industry that publishes only one half? That it sells through illusion, not through the honesty real consulting should be built on. To me, consulting means putting the whole truth on the table. Because anyone who is only ever shown one part of it keeps deciding blind.