AI Didn’t Fail. I Did.
A small, personal failure with AI reveals a strategic blind spot many leaders are still missing.
AI amplifies judgment; it does not replace it.
The true constraint in AI systems is human domain competence.
Leaders who confuse fluency with expertise will scale errors faster than insight.
It’s New Year’s Eve.
That strange, in-between hour where nothing urgent is happening, but everything important is quietly lining up in your head. It’s when leaders tend to reflect, not because someone told them to, but because the noise finally drops.
And in that quiet, I realised I ended the year by violating my own most basic rule about generative AI.
I asked ChatGPT to design a dog ramp.
A physical one. Timber. Angles. Load. Weather. Real-world consequences.
What followed wasn’t catastrophic. It was worse than that: it was plausible. Iteration after iteration that sounded right, felt thoughtful, and steadily drifted further from something I could actually build. The plan grew more complex. The certainty stayed high. My confidence quietly eroded.
Eventually, I stopped.
Not because the model failed, but because I finally recognised why it couldn’t succeed.
I had no domain knowledge.
And that’s the part many leaders are still underestimating.
Large language models don’t know when they’re wrong. They don’t reason the way humans do. They generate statistically coherent continuations. When the operator knows the domain, those continuations are extraordinary accelerants. When they don’t, the system produces something far more dangerous than error: convincing nonsense.
This isn’t a tooling problem. It’s a leadership problem.
We keep talking about “human-in-the-loop” as if it’s a compliance checkbox or a risk disclaimer. It’s neither. It’s a capability constraint. If the human in the loop cannot reliably evaluate the output, then the system is, functionally, unsupervised.
AI doesn’t democratise expertise. It rewards it.
In software, weak assumptions often surface quickly, tests fail, users complain, metrics dip. In physical systems, organisational systems, and strategic decisions, the feedback loop is slower and more expensive. By the time you realise something is wrong, the organisation has already aligned around it.
That’s the quiet danger CEOs need to sit with as the year turns.
Generative AI collapses distance between intent and action. That’s its power. But compression cuts both ways. It amplifies good judgment and bad judgment at the same speed. The bottleneck is no longer execution. It’s discernment.
I didn’t need a better prompt. I needed a carpenter.
And that’s the lesson.
As you head into the new year, reviewing strategy decks, approving pilots, green-lighting AI initiatives, ask a harder question than “Can we do this with AI?”
Ask: Who in this organisation can confidently say whether the output is right or wrong?
If the answer is “no one,” the model isn’t your risk. Your operating model is.
I’m ending the year with a dog who still jumps where a ramp should be, and a reminder that scales far beyond a backyard project:
AI doesn’t remove responsibility.
It concentrates it.
Happy New Year.
Use the quiet well.