260 Chicken McNuggets: Where AI Pilot Projects Hit the Wall
Just over a year ago, Klarna was the darling of every AI conference. The payment service provider announced that its AI assistant was doing the work of 700 customer service employees, and for a year, it hired virtually no one new. In 2025, there was a turnaround: The CEO admitted that they had focused too much on costs and ended up with “lower quality” as a result—now they’re hiring people again. The AI wasn’t stupid. It just wasn’t as good as the demo had promised. That only became apparent in real-world operations, with millions of real customers.
Klarna isn’t an isolated case. An IBM survey of 2,000 CEOs reveals a sobering statistic: Only one in four AI projects delivers the promised returns, and only 16 percent are ever rolled out company-wide. Nevertheless, everyone is rushing in—nearly two-thirds of executives invest before they’ve even understood the benefits, out of fear of being left behind.
Episode 2 focuses on the most costly stretch in any AI project: the gap between the demo that wows everyone and the Monday morning when the system is actually supposed to go live.
Key quotes from the episode
- “A demo is a story that ends with the demo.”
- “The demo thrived on the eighty-five percent. The rollout failed because of the fifteen.” — McDonald’s tested an AI ordering system in its drive-thru for three years. With an accuracy rate of around 85%, the final bill ended up including 260 Chicken McNuggets. The project was shut down in mid-2024.
- “The model isn’t the bottleneck. The path from the model to production is.” — MIT found that 95% of corporate pilot projects have no measurable impact on the profit and loss statement—and this is explicitly not due to the quality of the models.
- “The demo is cheap; operations are expensive.” — A good developer can build an impressive demo in two weeks. A system you can rely on takes months of work behind the scenes.
- “Show me a solution that’s been in production for a year. Tell me what broke along the way, and how you fixed it.” — The one question that separates salespeople from builders.
- “It’s not the most impressive solution that wins, but the one that’s still running on Tuesday.”
What It’s Really About
A demo runs on selected examples, with clean data, and with questions you’ve practiced answering. It shows the system’s best day, not the average—and our minds extrapolate from three flawless answers to three hundred. The missing 15 percent look like a small remnant that can be quickly cleaned up. In reality, they’re a hundred different edge cases, each with its own cause—and that’s exactly what the demo leaves out. That’s precisely where, usually very quietly, the expensive pilot project dies: the budget is gone, the enthusiasm too, and no one wants to be the one to pull the plug.
The good news: A few still make it into production, and they do things systematically differently. They build for the 15 percent instead of the 85 and look for value in the unassuming engine room rather than on stage. And you can tell from three simple questions whether you’re looking at an honest demo or a sales pitch. You’ll hear how that works—along with two anonymized case studies from our own projects that have been running continuously for years—in this episode.
