In the few short years since GenAI tools become widely available, we have regularly seen new articles about how AI can improve business models, and as a result, how many jobs will potentially be lost. The reality is not that simple, or that binary.
By Richard Firth, CEO, MIP Holdings
Right now, AI is being driven by commercial optimism, not engineering realism. Current large language models are extraordinary pattern predictors, but they are still statistical prediction engines. They predict the next token based on prior tokens. They do not reason, or understand, or form intent.
Despite all of the investment being put into various AI ecosystems, it is very unlikely that AGI (Artificial General Intelligence) is going to become a reality over the next few years. So, if we separate excitement from engineering, there is a very real possibility that the AI bubble, at least in its current form, will burst within the next few years.
A bubble bursting does not mean the technology disappears. The dot-com crash did not kill the internet, but it did clear the irrationality. The same will happen here.
Structural limits
As a software engineering company, MIP doesn’t experience AI as a consumer tool. We understand how it is built, the architecture, the scaling limits, the cost of training, the constraints of context, and the trade-offs behind every model decision. That level of technical literacy is still largely missing in many boardrooms where AI strategies are being defined today, and that gap matters because what we are seeing now is not the final form of AI, it is a phase.
The current generation of AI, particularly large language models (LLMs), is built on a relatively simple core idea: Predicting the next word based on what came before it. That prediction happens within a defined memory boundary, known as the context window, which limits how much information the model can consider at any given time.
Even at the upper end, this is finite. A model may handle hundreds of thousands, or in some cases close to a million tokens, but that still represents only a fraction of a real-world system, codebase, or enterprise environment, and that’s where the cracks begin to show. AI struggles to hold full systems in context. It can lose track of dependencies, generate inconsistent outputs, and, at times, produce confident but incorrect answers.
Techniques like prompt engineering and retrieval-augmented generation (RAG) help patch these gaps, but they don’t fundamentally solve them. At the same time, the industry is hitting diminishing returns from brute-force scaling.
Powerful, but not complete
The industry is being forced to confront the fact that bigger models are not the same as better intelligence. That realisation is already driving the next wave of innovation, which is being built on more efficient, specialised models instead of monolithic ones.
The limitations of today’s AI are actively shaping the breakthroughs of tomorrow, clearing the way for more grounded, more scalable, and ultimately more useful forms of AI. The real asset in the AI economy is not the model, it’s the data. Structured, operational, enterprise data that never lived on the open internet has enormous value and will lead to the dedicated models and enriched insights that AI is supposed to deliver, but fails to more often than not.
We are decades away from AI becoming a tool capable of replacing humans in the workplace, if that happens at all, so if you’re a software engineer or support specialist being told you’ll be replaced in 5 to 10 years, don’t worry. You’re not going to be replaced any time soon.
We are in the overestimation phase of the technology, and when the bubble bursts what remains will be a more grounded, practical wave of AI that delivers real value to a human workforce. A good example of the effects of overestimating the efficacy of AI is what happened at Salesforce. In late 2025, the company reportedly reduced thousands of roles under the assumption that AI agents could handle the workload. Executives later acknowledged they were too confident in the maturity of the technology as reliability issues surfaced and institutional knowledge was lost. The company began shifting back toward a human–AI hybrid model almost immediately.
This is becoming an increasingly important lesson for the market. The problem is not that AI lacks value, but rather that many organisations implemented AI expecting labour replacement and immediate cost reduction before the technology, economics, and governance models were mature enough to support those expectations at scale. In reality, AI introduces its own layers of cost, and for many organisations, AI is not replacing operational spend but adding another major operational layer on top of existing systems.
AI should be treated like any other powerful technology. Companies should understand its constraints, understand its cost, understand its failure modes, and keep humans in the loop where reasoning and accountability matter.
The bubble will burst. That’s the nature of every major technology cycle, but that reset is not a collapse, it’s a filter. Companies who don’t treat AI as a plug-and-play capability or a branding exercise will be the ones actually extracting value. When the hype fades, these organisations won’t be asking where they can use AI, they will already be using it in ways that are embedded, measurable, and commercially meaningful.