“Why don’t you just use the biggest AI?” The question comes up in almost every initial conversation, and it is a fair one. People mean the large general-purpose cloud models: they are impressive, immediately available, and the first bill is small.
In industry terminology, the largest language models at any given time are called frontier models, because they mark the current limits of what is possible. A language model is the tool that turns a question into a written response. Put bluntly: these models can do a little of everything, but know nothing from inside your business.
That is the key to the answer. And “just” is the part of the question that does not hold up.
World knowledge and company knowledge
A large general-purpose model has world knowledge. It knows how a centrifugal pump works, which standards apply to pressure vessels, and the damage cavitation causes. That is a great deal, and it is impressive.
What it does not know is which pump is on Line 3 at the North Plant, when it was last serviced, what the manufacturer wrote about it in its 2019 approval, and what exception the workshop manager recorded in an email three years later.
That is company knowledge. It is in a manual on a file server. In a drawing nobody has opened since 2014. In an inspection report someone scanned. And a good deal of it is written nowhere, but passed on at shift handover.
No language model has ever seen any of it. A frontier model grew up on generally accessible material: standards, textbooks, forums and manufacturers’ websites. The 2019 approval was not among them. Neither was the workshop manager’s email. And information passed on at shift handover is not in any dataset in the world.
The problem, then, is not a lack of computing power. It is a lack of access. No increase in model size changes that.
The quality of the answer is not determined by the size of the model. It is determined by access to the company’s knowledge.
And that is how it should stay. There is exactly one way for an external model to learn which pump is on Line 3: someone would have to tell it. That someone would be you. Or someone in your company who was in a hurry.
Why a language model is needed at all
It receives the retrieved passages and uses them to formulate an answer, with a source reference for each statement. A narrowly defined task.
Finding and formulating are two different things. Finding is the hard part: who may see which document, what a term means in this business, how to retrieve a scan from 1998, or an English manual when the question is asked in German. Formulating is the easier part, and that is what the model handles.
If the hard part has not been solved, a larger model gives you a more elegantly worded wrong answer.
What actually determines system sizing
When sizing such an installation, people usually start with the number of users. That is the wrong measure. What matters is how many questions are asked at the same time, which is a very different number. You know how high it is only after measuring it. Before sizing the system, the question is therefore not how many people work in the company, but when many of them ask questions simultaneously.
There is a second measure alongside it, and the two are often confused. Answering a question is routine operation: it happens every time someone wants to know something, and it is quick. Ingesting the documents happens once, when the knowledge base is built, and afterwards only for newly added material.
How long that first ingestion takes depends almost entirely on the types of sources. A text document is read in a fraction of a second. A scanned page with handwriting takes many times longer, not because the process is slower, but because more happens. The page is divided into lines. Each line is read individually by a model that understands images. Each line receives a score indicating the confidence of recognition. And every line with a score that is too low is cropped out and presented to a person who decides what it says.
Anyone connecting file repositories and databases will notice little of the ingestion process. Anyone connecting a paper archive accumulated over many years needs to plan a separate undertaking for it.
A larger language model changes neither of these things. It may produce more polished wording. It does not make the answer faster, and it knows no more about your equipment.
It does not fix poor answers either. If answers are unusable, the cause is almost always upstream of the model: in what it is given, and in what form. Replacing the model then means replacing precisely the part that was not the problem.
The second part of the question
That leaves the cloud service, the more convenient half of the question. It is called SaaS, Software as a Service: the application runs at the provider, many customers share the same installation, and you rent access rather than operating anything yourself. No hardware in the server room, no operations to manage, immediately available. Concerns are usually dealt with by ticking a box: your data will not be used for training.
That box is not meaningless. It is a commitment about what the provider does with your documents. Your documents have nevertheless left your company.
A commitment can change with new terms, a new pricing plan or a new owner. And it cannot be independently checked from outside. You can demonstrate that something never left. You cannot demonstrate that it was deleted.
This is the important distinction: the service is replaceable; the knowledge collection is not. A language model can be swapped. In NXTGN AI CORE, that takes one configuration line, because the language model, embedding and image understanding sit behind interfaces rather than being hard-wired into the core. Changing the embedding model also requires re-indexing; changing the language model does not. A collection, by contrast, grows over years, and the decision to send it outside the company is made once.
That is why NXTGN AI CORE is built so that nothing leaves the customer’s premises. An image that needs describing is sent within the request itself, never as a link to an external computer. Internal services listen locally only. These are not grand architectural gestures. They are implementation details, and it is precisely details like these that make the difference.
Where a frontier model is worthwhile
I do not want to be misunderstood. There are tasks for which the largest available model is the better choice.
Whenever the answer needs to come from the model itself, without a supplied source. For long chains of reasoning. For less common languages. Someone who wants a draft written is well served by the largest model. Someone who wants to know what their own manual says is not.
These cases exist, and they can be counted. They do not change the rule; they define its boundary.
My conclusion
Much of the quick approach makes sense. The large service is immediately available, nobody has to operate anything, the first bill is manageable, and the result looks good in a demo.
But it does not answer the question asked at seven in the morning on the production line. That question is not “What do you know about centrifugal pumps?” It is “What do our own records say, and am I allowed to see it?” That needs no larger model. It needs the company’s own knowledge, the permissions for it, and a source reference beneath every statement.
Connect sources, clarify permissions, check scans. That is the less convenient work, and it never looks impressive in a demo.
In a few years, nobody will be talking about today’s biggest and best AI model. They will still be talking about the data sources someone connects now.