When ChatGPT launched in November 2022, the reaction was immediate and visceral: this works. For the first time, millions of people experienced AI not as a distant promise, but as something useful, intuitive, and even with its flaws, astonishingly capable.
That instinct was correct. The conclusion that followed was not. Because what works brilliantly for an individual at a keyboard has proven surprisingly ineffective inside an organization.
Two years later, after billions in investment, countless pilots, and an endless stream of “copilots,” a different reality is emerging: generative AI is exceptional at producing language. But companies do not run on language: they run on memory, context, feedback, and constraints. That’s the gap. And that’s why so many enterprise AI initiatives are quietly failing.
High adoption, low impact… and a growing sense of déjà vu
This is not a story about a technology that failed to gain traction. It’s the opposite.
A widely cited MIT-backed analysis found that around 95% of enterprise generative AI pilots fail to deliver meaningful results, with only about 5% making it to sustained production. Other coverage of the same findings points to the same pattern: massive experimentation, minimal transformation.
And the explanation is telling: the problem isn’t enthusiasm, or even capability: it’s that the tools don’t translate into real, operational change.
This is not an adoption problem. It’s an architecture problem.
The uncomfortable paradox: everyone uses AI, but nothing changes
Inside most companies today, two realities coexist: on one side, employees use tools like ChatGPT constantly. They draft, summarize, ideate, and accelerate their work in ways that feel natural and effective.
On the other, official enterprise AI initiatives struggle to scale beyond carefully controlled pilots.
The same MIT-related analysis describes a widening “learning gap”: individuals quickly find value, but organizations fail to integrate that value into workflows that matter. The result is something close to “shadow AI”: people use what works, while companies invest in what doesn’t.
That’s not resistance to change.
That’s a signal.
The core mistake: treating a language model like an operating system
Most explanations for this failure focus on execution: bad data, unclear use cases, lack of training. All true. All secondary.
The real issue is simpler and far more fundamental: large language models are designed to predict text. That’s it. Everything else, from reasoning, to summarization, conversation, etc. is an emergent property of that capability.
But companies do not operate as sequences of text. They operate as evolving systems with state, memory, dependencies, incentives, and constraints.
This is the mismatch.
As I’ve argued before, this is AI’s core architectural flaw: LLMs do not “see” the world. They do not maintain persistent state. They do not learn from real-world feedback unless explicitly engineered to do so.
They generate convincing language about reality. They do not operate within it.
You can’t run a company on predictions of words
This leads to a pattern that should feel familiar.
Ask an LLM to:
- “Increase my sales”
- “Design a go-to-market strategy”
- “Improve team performance”
And you will get an answer. Often a very good one. A structured, articulate and persuasive answer. And almost entirely disconnected from the actual system it is supposed to influence.
Because an LLM cannot track a pipeline, manage incentives, integrate CRM data, or adapt based on outcomes. It can describe a strategy. But it cannot execute one.
The MIT findings reinforce this point: generative AI tools are effective for flexible, individual tasks, but break down in enterprise contexts where adaptation, learning, and integration are required.
In other words: an LLM can write the memo. But it cannot run the company.
Throwing more compute at the problem won’t fix it
The industry’s response so far has been predictable: build bigger models, deploy more infrastructure, scale everything. But scale does not fix a design flaw. If a system lacks grounding in reality, more parameters will not give it grounding. If it lacks memory, more tokens will not give it memory. If it lacks feedback loops, more data centers will not create them.
Scale amplifies what exists. It does not create what’s missing. And what’s missing here is not more language. It’s more world.
The next layer won’t be about better answers
The next phase of enterprise AI will not be defined by better chat interfaces or more powerful LLMs. It will be defined by something else entirely: systems that can maintain state, integrate into workflows, learn from outcomes, and operate under constraints.
Systems that don’t just generate text, but act within real environments. This is why the future of AI in companies will not be built on LLMs alone, but on architectures that embed them within richer models of reality.
Or, as I’ve argued in previous work, why world models are likely to become a foundational capability rather than a niche concept.
Saying what many already know… but rarely say
If this feels obvious, it’s because many people inside organizations already see it: they’ve run the pilots. They’ve seen the demos. They’ve experienced the gap. But saying it out loud is still uncomfortable.
There is too much momentum, too much investment, and too much narrative built around the idea that scaling LLMs will eventually solve everything. It won’t.
The emperor is not just underdressed. He is wearing the wrong clothes entirely.
The real opportunity
This is not the end of enterprise AI: it is the end of a misconception. Language models are not enterprise architecture: they are an interface layer. A powerful one, but insufficient on its own.
The companies that understand this first will not just deploy AI better: they will build something fundamentally different.
And when that happens, it will feel, once again, like magic.
But this time, it won’t be an illusion.
