In a previous article, I broke down the difference between AI, machine learning, and generative AI. That distinction matters, but it also leads to a more important question: If we understand what these technologies are… why do so many AI projects still struggle to deliver value?
In most cases, the failure has very little to do with the AI itself. The assumption that the model is the hard part comes from a common belief that AI projects are difficult because the technology is complex. That used to be true, but today, it’s far less of a barrier.
You can access powerful models through APIs from companies like OpenAI, Anthropic, or Google DeepMind without needing to build anything from scratch. Other open-source models are improving rapidly too. Tooling has matured, and infrastructure is more accessible than it has ever been.
In other words, the model is no longer the bottleneck, and yet projects still stall and fall over.
Where things actually go wrong
The problems tend to show up much earlier, and much more quietly. It often starts with a vague objective.
“We want to use AI”
That often sounds like a strategy, but it isn’t a problem definition. Without a clear use case, it’s difficult to measure success, and even harder to design something that delivers it.
Then there’s data. Most organisations don’t have a clean, well-structured dataset ready to support machine learning. Data is fragmented across systems, inconsistent in format, or simply not captured in a way that makes it usable. Instead of addressing that, projects move forward anyway, layering AI on top of unstable foundations.
Integration is another friction point. Even when a model performs well in isolation, it still needs to fit into existing workflows. It needs to connect to systems, be trusted by users, and produce outputs that can actually be acted on. Without that, it remains a proof of concept rather than something that delivers operational value.
None of these issues are particularly visible at the start. But they compound quickly.
The pilot trap
A pattern that shows up repeatedly is the “successful pilot that goes nowhere”. A team builds a working prototype. It demonstrates potential. There’s interest .......... and then… nothing.
Scaling requires more than a working model. It requires:
- Reliable data pipelines
- Governance and ownership
- Integration into business processes
- Ongoing monitoring and iteration
Without those, the project stalls between experimentation and production. This is one of the reasons organisations like Gartner have consistently reported that a large proportion of AI initiatives fail to move beyond pilot stages.
The challenge isn’t proving that something can work. It’s making it work consistently, reliably, and at scale.
Overengineering too early
Another common issue is jumping straight to the most advanced solution available. Whether it’s retrieval-augmented generation (RAG), multi-agent systems, or other complex architectures, these can all be powerful when used in the right context. But they also introduce cost, complexity, and maintenance overhead.
In many cases, the underlying problem could have been solved with something far simpler, maybe a structured workflow, a rules-based system, or a lighter, more basic model. Starting with complexity doesn’t just increase risk, it makes it harder to understand whether the solution is actually delivering value.
Even when the technology works, and the data is in place, there’s still one more challenge: people.
If users don’t trust the output, they won’t use it. If it doesn’t fit into how they already work, they’ll bypass it. And if it creates more friction than it removes, it won’t last.
AI doesn’t replace decision-making, it changes how decisions are made. That shift needs to be understood and managed, otherwise the solution never becomes part of the business.
What this looks like in practice
Across most projects, the same pattern emerges: the focus is placed on the AI when the actual constraint sits somewhere else.
The constraint could be data quality, process design, or clarity of the project’s objective, but the model itself is rarely the reason things fail.
That doesn’t mean AI is simple. It means the difficulty has moved.
A more grounded starting point
A more effective approach is to work backwards from the outcome:
- Start with a clearly defined problem: your business need.
- Understand the data that supports it: do you have that data?
- Look at how the solution will be used, not just how it will be built.
- Then decide whether AI is actually required and, if it is, which type.
That sequence might sound obvious, but it’s also where most projects diverge.
Final thought
AI has become easier to access, easier to implement, and easier to experiment with, but delivering value from it hasn’t become easier. It’s just changed shape.
The hard part is no longer building the model. It’s everything around it.
Next article: In the next piece, we’ll go deeper into one of the most common failure points: data. Why “garbage in, garbage out” becomes significantly more dangerous in an AI context, and why most organisations underestimate how much work sits there before any model is even considered.