Interesting article – it notes that roughly 25% of corporate AI projects actually deliver ROI:

Fortune, 2025/05/09: "Klarna plans to hire humans again, as new landmark survey reveals most AI projects fail to deliver"

I'd like to say I'm surprised by this figure, but I'm not. It feels a bit on the high side, really.

It's easy to blame this all on Corporate FOMO™ and executives rushing headlong into AI. I'm sure that played a role in this dismal success rate. But there's more to it than that.

Of the 75% of use cases that didn't pan out, we can ask which were:

1/ Good ideas that just didn't pan out. (This is a hard, yet often-overlooked aspect of any kind of data/ML/AI project. It happens. And it's no one's fault. You have to take the risk if you want a shot at the reward.)

2/ Ideas that people knew were long shots, but decided to try anyway. (These are great. So long as you acknowledge in advance that these are Unlikely To Pan Out But Worth Gold If They Do, and fund them accordingly, then you're fine.)

3/ Ideas that were long shots, but people did not know this was the case. (Sometimes this happens when a project is oversold.)

4/ The aforementioned Corporate FOMO™ and other Let's Just Do This To Do This projects. (You know the drill.)

When evaluating success of an AI project, it's important to know which of those four best describes it.

We can also ask two questions about the remaining 25% of use cases that made it through:

How many genuinely work, and how many "work for now, by coincidence?" In other words: how many of those projects carry a lot of unacknowledged exposure to downside risks?

("Risk" is fine. "Risk we don't see" is not fine.)

Of those that work, what kind of ROI do they bring? Is it more "meh" or "absolutely revolutionized our business?" Remember that success is not always a matter of yes/no; degrees of success matter.

(Having trouble sorting out your use cases for ML and genAI? Reach out. Let's work together to find the best ways to put this technology to good use)