AI, Honestly
The Last 10 Percent Is Where AI Projects Die
Every AI project that lands on my desk for a rescue has the same fingerprints. The pilot was a triumph. The demo drew applause. Then, somewhere between the conference room and real customers, the whole thing quietly stopped moving. Nobody killed it. It just got renamed "phase two" and slid off the roadmap.
After 35+ years building software, systems, and security, I can tell you the failure was never the model. It was the last 10 percent. And the last 10 percent is where AI projects actually live or die.
The 90 percent that demos beautifully
Here is the uncomfortable truth about AI hype versus reality: the first 90 percent of almost any hard problem is the part that shows well. Clean inputs, a curated dataset, a narrow happy path, a friendly audience who wants it to work. Modern AI tools have made that first 90 percent faster to build and far easier to fake than ever before. You can stand up something that looks like magic in an afternoon.
The problem is that the demo is measuring the easy part. Production does not run on curated inputs. It runs on whatever your customers actually type at 2am, the contradictory records, the edge cases nobody scoped, the exception that shows up eleven times a week and was never written down. That last 10 percent is not a rounding error. It is where the system either ships, scales, and survives a Monday morning, or it does not.
Why the last 10 percent is the whole game
I once built an AI system for sports video that had to find the marketable moments in a live football match using facial recognition. In the demo, on hand-picked clips, it was stunning. Then we pointed it at real Premier League footage and it fell apart. At one point it confidently identified a player who had been retired for two decades. More AI was not the answer. The fix was a constraint borrowed from experience: stop trying to recognize every face on earth and limit the model to the 22 players actually on the pitch. Accuracy went through the roof.
That is the last 10 percent in a nutshell. It is rarely solved by a better prompt or a bigger model. It is solved by judgment, by narrowing scope, and by someone who has shipped the ugly part before and knows exactly where it bites. The teams that win at AI are not the ones with the flashiest pilot. They are the ones who scoped the ugly 10 percent before they wrote a single line of code.
Where does AI fit? Start by pricing the hard part
When a founder asks me "where does AI fit in our business," the honest first move is not to point at the exciting use case. It is to ask what the last 10 percent of that use case looks like. Can we see a path through the messy inputs? Do we know how the model behaves on the traffic we cannot control? Is there a constraint that makes the problem tractable, the way limiting the model to 22 players did? If we cannot answer those before we start, we do not start. I make that exact call in my own three companies every week.
This is where a fractional CTO or AI advisor earns their keep, and it has almost nothing to do with knowing the latest model. The value is in reading the gap between a good demo and a system that survives production, then telling you honestly whether that gap is a speed bump or a wall. A good AI consultant should be able to save you a six-figure "phase two" by pressure-testing the idea for the price of a conversation.
The honest version of the question
AI is not overhyped. Specific, unscoped promises about AI are. The winners in this next stretch will not be the companies that fell in love with what AI could do in a controlled room. They will be the ones who asked the boring, disciplined question first: does this hold up on the inputs we cannot control? Answer that well, and the last 10 percent stops being where your project dies. It becomes the reason it lasts.
If you have a pilot stuck in phase two, or an AI idea you want stress-tested before you spend real money on it, let's find out whether the last 10 percent is a speed bump or a wall. I take a few clients at a time. Book your free 15-minute AI Fit Call →
Stephen Bickle · Jul 26, 2026