Grapine's founder has been building AI systems for enterprise organisations since before most of the current conversation existed. Not in the sense of reading about it, advising on it, or building prototypes. In the sense of shipping production AI into large organisations and living with the consequences when it did and didn't work.
That experience doesn't produce cynicism about the current moment. The capability jump of the last few years is real, genuinely significant in a way that previous waves weren't. The things that are now possible would not have been possible five years ago, and the difference matters.
But experience does produce something useful: the ability to recognise what hasn't changed. And several things have not changed across 25 years of AI waves, despite the technology changing dramatically underneath them.
These are the things the hype cycle consistently forgets.
The demo has always worked
Every AI wave produces impressive demos. The 2024 version is more impressive than the 2015 version, which was more impressive than the 2008 version. But in every case, the demo worked.
Production is different from the demo in ways that are predictable, specific, and consistently underestimated. In a demo, the inputs are clean, the edge cases are handled, and the failure modes are not shown. In production, users send inputs the system wasn't designed for, edge cases arrive on the first day, and failure modes surface in ways that are obvious in hindsight and invisible in advance.
This is not a criticism of AI. It is a description of how every technology behaves when it moves from controlled demonstration to real-world deployment. The gap between demo performance and production performance has always existed. It still exists. The businesses that close that gap deliberately, by testing adversarially, by building verification into workflows, by designing for failure modes before they occur, are the ones that get the value. The businesses that ship the demo into production are the ones that end up explaining why it didn't work.
The people problem has always outlasted the technology problem
In every AI implementation we have been part of, the technology was ready before the organisation was. The model worked. The integration was built. The tool was functional. And then it was underused, or misused, or quietly abandoned, not because the technology failed but because the people around it hadn't changed how they worked.
This pattern predates large language models by decades. It is not a criticism of AI, it is a description of how organisations absorb new capabilities. People change more slowly than technology. Workflows built over years don't restructure because a new tool becomes available. Trust in a new system is earned slowly, through demonstrated reliability, and lost quickly, through a single conspicuous failure.
The AI implementations that have generated real value across every wave share a common characteristic: they were designed with adoption as a constraint, not an afterthought. The question wasn't just "can AI do this?" It was "Will the people who need to use this actually use it, and what does the system need to look like for that to be true?"
That question is still the right one to start with.
The data has never been as ready as assumed
Every AI project begins with an assumption about the data. It will be accessible, it will be clean, it will be in a usable format, and it will cover the scope of the problem.
It is never all of these things.
Data readiness is the most consistent source of delays and cost overruns in AI implementation, and it has been for as long as AI has been implemented in real organisations. Not because organisations are careless with their data, but because data that is adequate for human use is frequently not adequate for machine use. Formats that a person can interpret require transformation before a model can use them. Records that are complete enough for reporting have gaps that matter for training or inference. Systems that were built to store information were not built to export it in the form an AI integration needs.
This hasn't changed. What has changed is that the current generation of tools is more tolerant of messy data than previous generations were. That tolerance is not unlimited. The businesses that assess their data honestly at the start, not optimistically, save significant time and money compared to those that discover the problem partway through a build.
AI has never replaced judgment. It has always changed what judgment is applied to.
Every wave of AI arrives with some version of the same claim: this one will automate the things that require expertise. This one is different from the previous waves because it can reason, not just pattern-match.
There is always something to this claim. The capability boundary does move. Things that required human expertise two decades ago don't require it now.
But the boundary moves more slowly than the announcements suggest, and in a specific direction: AI consistently absorbs the mechanical and the routine before it absorbs the judgment-intensive. The things that required expertise but were actually formulaic, that looked like judgment but were really pattern application, get absorbed first. The things that genuinely require weighing ambiguous information in context, with accountability for the outcome, take much longer.
What changes with each wave is what the judgment gets applied to. Professionals who used to spend 60% of their time on research and 40% on advice now spend more time on advice. The proportion shifts. The judgment doesn't disappear; it gets more room.
This is still true. The current generation of AI is genuinely better at the research and synthesis end of professional work than anything that preceded it. That frees time for judgment-intensive work. It doesn't eliminate the need for it.
The bottleneck always shifts
Early in any AI project, the bottleneck is capability: can the technology do what we need it to do?
After that question is answered, and increasingly, the answer is yes, the bottleneck shifts to integration: can we connect this to our actual systems and workflows?
After that, it shifts to adoption: will the people who need to use this actually use it?
And after that, it shifts to trust: does the organisation trust the output enough to act on it without excessive verification overhead?
Every AI wave has worked through these bottlenecks. The current wave is working through them too, faster than previous waves because the capability starting point is higher. But the sequence is the same. Understanding which bottleneck you're currently facing is more useful than assuming the technology is the problem when adoption is the real issue, or assuming adoption is the problem when the integration isn't actually solid.
What the current moment actually is
The hype cycle around AI is real, and it obscures something important: the underlying capability change is also real, and it's larger than previous waves.
What makes this wave different from its predecessors is not that it replaced judgment; it hasn't. It is that it made a specific category of work, the research, the synthesis, the first draft, the structured extraction of information, dramatically cheaper and faster to do. Work that was bounded by the speed at which humans could read, process, and write is now bounded by different things.
That is a genuine change. The businesses getting the most value from it are not the ones chasing the most impressive demos. They are the ones who identified a specific problem, defined what good output looked like, built a system that produces it reliably, and designed around the failure modes that appear when real users arrive.
That is how every AI wave has created value. The specific technology has changed. The approach hasn't.
Grapine is built on this perspective by a team that has shipped production AI across 25 waves and knows the difference between capability and deployment. If you want that grounding applied to your business, the AI Audit is where we start.