Ask any business what it actually does and the answer is always specific. A quarrying company extracts and processes rock. A conservation organisation restores damaged peatlands. A fashion retailer curates and sells clothing. The answer is never "we send emails" or "we use spreadsheets" or "we have meetings." Those are the generic operations that every business shares. They are not why the business exists.
This distinction, between what a business does and the common operations surrounding it, has always mattered. Adam Smith observed in the eighteenth century that specialisation was the engine of productivity. Workers focusing on narrow tasks outperformed generalists. Economies grew by dividing labour into ever finer slices. The principle has only deepened since: modern businesses succeed by filling specific niches, doing particular things better than the alternatives.
Artificial intelligence is now arriving in force, and it will reshape how businesses operate. But it will not change this fundamental dynamic. If anything, it will sharpen it.
The Coming Ubiquity
General purpose AI, particularly large language models, will be adopted by virtually every business. This is already happening. These tools handle communication, documentation, analysis, and routine decision support well enough that ignoring them will become a competitive disadvantage.
But this ubiquity has a consequence. When every business uses the same tools for the same generic tasks, those tools become infrastructure, like email or spreadsheets. Essential, but not what sets anyone apart. No company gains competitive advantage from having access to email. Soon, no company will gain competitive advantage simply from having access to LLMs.
There is also a more fundamental limitation. These tools excel at tasks that are common across organisations precisely because they are trained to be generalists. They can summarise documents for a quarrying company just as well as for a fashion retailer, because summarising documents is the same task in both contexts. The moment a task becomes specific to what the business actually does, general tools begin to struggle.
A quarrying company does not just need help with emails. It needs to optimise blast patterns based on geological conditions, monitor crusher efficiency in real time, and match production to market demand for different aggregate grades. A conservation organisation does not just need help writing reports. It needs to map erosion across thousands of hectares of remote terrain from aerial imagery, then plan restoration interventions accordingly. A mining company searching for ore deposits needs to analyse sparse 3D geological data and predict where concentrations might lie in unexplored areas.
These are the core functions, the actual niche each business exists to fill. General AI has little to offer here, not because it has failed, but because generality and specificity are opposites. A system trained to do everything adequately cannot also be expert in the particular.
Building for the Specific
This is where specialised AI becomes necessary.
We have built systems across these domains. For a major quarrying operation, a computer vision system that monitors rock size distribution on crusher conveyors, detecting hidden downtime that amounted to 5% to 10% of operational hours. For peatland restoration, a transformer-based network that maps erosion from aerial imagery and 3D depth data, dramatically accelerating what was previously slow and expensive manual assessment. For mineral exploration, a 3D scene completion model that predicts ore concentrations in unexplored areas, achieving accuracy levels that earned recognition in an international competition.
Each system works because it was designed around a specific problem: trained on domain-relevant data, built using architectures suited to the task, optimised for the metrics that actually matter in that context. Often this means smaller, more efficient models rather than larger ones. Specialisation is not about adding complexity. It is about removing everything that does not serve the specific goal.
The results are not incremental improvements. A demand forecasting model built around a retailer's specific data, product categories, promotional patterns, and supplier constraints can halve forecasting error compared to generic approaches. That is not a marginal gain. It is the difference between a system that informs decisions and one that gets ignored.
The Logic Ahead
As AI moves from novelty to infrastructure, the pattern will become clearer. General tools will handle what is general: the operational tasks every organisation shares. These will become table stakes, like internet access or accounting software. Necessary, but not distinctive.
The functions that define a business, the specific niche it exists to fill, will require AI built for that purpose. Just as businesses differentiate by specialising, the AI that enables their core operations will specialise too. This is not a prediction about technology. It is simply the economics of specialisation applied to a new domain.
At New Gradient, we build bespoke AI systems designed around the specific processes and data of each client, from industrial optimisation and environmental monitoring to demand forecasting and geological analysis. If you are considering how specialised AI could strengthen what your business actually does, we would be glad to discuss it.