Is this the year we unravel the AI conundrum?
We’ve been working with machine learning for nearly a decade. For most of that time, the advice was straightforward: ML is a set of techniques, each suited to specific problems, with reasonably clear ways to assess where it helps and where it doesn’t.
That clarity has become harder to maintain since the field rebranded as AI.
Where ML was targeted and measurable, AI has been positioned as a cure-all. The superficially impressive outputs of large language models like ChatGPT, Claude and Gemini seemed to justify the hype, and the genuine successes of ML were quietly folded into AI’s broader story. For a while, the belief held.
It’s proving harder to sustain. As the limitations of generative AI become more visible, we find ourselves returning to the same questions we asked before: what problem are we actually solving, and is this the right tool for it?
Some use cases genuinely stand up. Document summarisation, broad research and data transformation are good examples: tasks that are difficult to automate with traditional software, where AI delivers real value. But there are probably fewer strong use cases than the market has suggested.
Our view is simple: you can’t make good business decisions about AI without understanding what “AI” actually refers to in a given context. The intelligence isn’t in the model. It’s in how well the output fits the problem you’re trying to solve.
