I have shipped statistical models, machine learning, production NLP, computer vision, and now agentic AI. The pattern across all of it is the part worth paying attention to.
I started building statistical models more than thirty years ago. Then came machine-learning-based predictive models. Then natural language processing in production, which in its early days meant fine-tuning the system with human analysts because there was no other way. Then computer vision at scale on satellite imagery. Now LLM-based agents. Five waves, each one announced as a break with everything before it.
What I learned watching each wave arrive is that the technology is the part that changes and the discipline is the part that does not. Every wave produced impressive demonstrations. Every wave produced a graveyard of deployments that worked in the lab and collapsed in production. The deployments that survived shared the same traits across all five waves, and none of those traits were about the model.
What survived, every time
The systems that lasted had three things. First, defensible measurement. Someone could explain why the output was trustworthy to an executive, a client, or a regulator, and the explanation held up. Second, a workflow that actually changed. The technology was embedded in how people did their jobs, not bolted onto the side. Third, an owner. A specific person was accountable for the result after the consultants and the vendors left.
The systems that collapsed were missing at least one of those, no matter how good the model was. This is why I am skeptical of anyone whose AI credibility starts in the last two or three years. They have seen one wave. They have not seen what happens to the second-best demo six months after launch.
Why this matters for your investment
If you are spending on AI right now, the relevant question is not whether the model is impressive. It is whether your organization can operate the result. That question is older than the current wave and it will outlast it. The technology will keep changing. The discipline of making it survive contact with real data and real users will not.