AI is brilliant at first encounters.
Give it a product and it will read, infer and describe. Give it ten thousand products, arriving through different suppliers, written in different vocabularies, and ask it to make the same decision every time, and brilliance begins to look rather less like discipline.
The problem is not intelligence. It is memory, consistency and governance.
A good answer is not yet a system
An AI tool may look at a hand-thrown ceramic vessel and suggest:
Category: Home décor
Product type: Vase
Material: Stoneware
Colour: Sea green
Useful. Perhaps even correct.
But a catalogue does not contain one product. The next maker calls the same form a bud vase. Another calls it a stem vessel. A third lists it as decorative pottery. The glaze is sea green, celadon, sage and ocean, depending on who entered it.
The question is no longer whether AI can understand each line. It probably can.
The question is whether all those understandings become one dependable commercial language.
The prompt becomes a policy manual
The obvious response is to improve the prompt.
Tell the model that bud vase and stem vessel should be related. Tell it that celadon should not automatically become sage. Explain when reactive describes a glaze and when it is merely promotional copy. Add the permitted categories. Add exceptions. Add formatting rules. Add examples of earlier mistakes.
Soon the prompt is carrying the taxonomy, the training notes, the quality policy and the catalogue's institutional memory.
That is not necessarily a failure of prompting. It is evidence that the problem has become larger than a prompt.
Corrections disappear into the next call
A merchant corrects a product once. What should happen next?
In a dependable system, the correction should become reusable knowledge. Similar products should benefit from it. The decision should be visible, reviewable and reversible.
In a pure generation workflow, the next product may simply be sent through another fresh inference. The correction survives only if it is added to context, stored elsewhere or converted into a rule.
At that moment, the system has quietly admitted what the catalogue needed all along: structure.
Variation becomes expensive to review
AI outputs do not have to be absurd to create work. They only need to vary at the edges.
If 90 per cent of suggestions are obvious and 10 per cent contain commercially important mistakes, a person still has to find the 10 per cent. Review becomes its own form of labour.
The more specialist the catalogue, the more those edges matter. A generic category may be broadly defensible and still be wrong for how the retailer merchandises the product.
Repeated inference creates repeated cost
If every product requires another model call, scale carries a variable cost. The expense may be modest per item and still become meaningful across large catalogues, reprocessing, new imports and repeated corrections.
More importantly, the business keeps paying to rediscover decisions it has already made.
Use AI for the work that needs AI
AI is exceptionally useful for interpreting messy language, detecting relationships and accelerating calibration.
But once a merchant has confirmed that wine, in this catalogue, belongs under Burgundy, the system does not need inspiration. It needs obedience. Understanding a product and applying that decision consistently are different tasks.
Semantikal harnesses AI to understand the catalogue, then turns approved understanding into logical structure. The AI interprets. The merchant decides. The logic remembers.
That is how intelligence becomes infrastructure.