Businesses are moving quickly from experimenting with artificial intelligence to putting it into real workflows. The UK government says business adoption has more than doubled since December 2023, reaching 26% in March 2026. As AI moves deeper into everyday business operations, the more important question is becoming how reliably these systems perform when the details have commercial consequences.
AI can produce an answer that looks correct while missing the detail that actually matters. I saw this first-hand while building LigaxAI. In early testing, a leading general-purpose model could correctly identify an image as showing braids. On the surface, that looked like a successful result. It could not reliably distinguish between knotless braids and box braids, however. To the model, these were relatively minor visual differences. For a salon, these are different services with different prices, booking times and installation requirements. The model had recognised the subject, but missed the information the business needed.
That experience made me think differently about how businesses measure AI performance. Accuracy does not have a single meaning. It depends on the task, the information required and the consequences of getting the answer wrong.
A general-purpose model is trained on an enormous amount of information. That breadth is one of its greatest strengths. The world’s data does not represent every distinction equally, however. Some things are photographed more often, labelled more consistently and discussed more widely than others. Highly specific professional knowledge can therefore be poorly represented, even when that knowledge is commercially important.
This becomes particularly important when we consider how AI learns. The labels used to train an image recognition model help determine what distinctions it learns to recognise. A generalist looking at a hairstyle and a professional hairstylist may assign different labels to the same image because they understand the distinctions differently. A stylist knows which visual features determine the technique, the service and ultimately the price.
For me, that is one of the most important lessons from building LigaxAI. The label is as important as the image. Domain-specific data is not simply more data. Its value comes from the expertise behind it and the meaning that has been verified by people who understand the field.
The timing matters because businesses are moving AI beyond experimentation. The latest UK Business Data Survey found that 41% of businesses handling digitised data reported using AI in 2025–26, while only 21% of AI-using businesses said their tools were integrated into existing business systems. The figures suggest that adoption is moving ahead of full integration, leaving many organisations still working out how AI should fit into established processes.
That makes the definition of “good enough” increasingly important.
If an AI tool produces an awkward summary, the consequence may be negligible. If it misclassifies a product, misprices a service, makes the wrong compliance recommendation or gives a customer incorrect information, the consequences can be very different. The more consequential the decision, the more important it becomes to understand exactly how the system performs on the task it has been given.
Businesses therefore need to consider one fundamental question before deploying AI. What is the cost of a wrong answer? If a mistake is inconvenient but easily corrected, a general-purpose model may be perfectly suitable. If an error can lead to lost revenue, regulatory exposure, customer dissatisfaction or operational disruption, businesses need to test the model against the specific data and decisions it will encounter in production.
This is also where specialised AI is often misunderstood. There is an assumption that building a specialised system must automatically be more expensive or slower than configuring an existing general-purpose model. The more useful comparison is between the investment required to achieve the necessary level of performance and the ongoing cost of correcting errors.
A business can spend months refining prompts, reviewing outputs and adding safeguards while continuing to encounter the same failure modes. For a narrow, high-value task, building around specialist data and expertise can provide a more direct route to the performance the business requires.
The wider AI industry is beginning to confront this reality as businesses move from pilots into deployment. The UK government has identified skills gaps, data access, governance uncertainty and the difficulty of moving from pilots to sustained deployment as barriers to wider AI adoption. These challenges all point towards the same practical requirement. Businesses need to understand the specific task, define what a successful outcome looks like and measure performance against the conditions in which the system will actually operate.
That requires better data and better testing. It also requires people who understand the industry well enough to determine which distinctions matter. In specialist environments, domain expertise can turn a large volume of information into useful training data by giving the system a clearer understanding of what it is actually being asked to recognise.
The AI industry has become very good at measuring general capability. Businesses now need to become better at measuring whether that capability translates into the decisions they depend on.
The most valuable AI capability for a business may be knowing what matters. And as AI becomes more deeply embedded in the decisions businesses make every day, the difference between being broadly right and precisely right will become increasingly difficult to ignore.
