Definition
An LLM predicts useful continuations of language based on patterns learned during training. It can be given instructions and context at the moment you use it, but it does not automatically know anything about your business.
How an LLM works, in terms that matter for a business decision
An LLM was trained by reading an enormous amount of text and learning which words tend to follow which. Given a prompt, it produces the continuation that best fits the patterns it learned. That is genuinely the mechanism. There is no lookup table of facts inside it, and no database of your company.
Three consequences follow, and every one of them shows up in real projects:
It does not know your business. Not your prices, your policies, your customers, or what happened last Tuesday. Anything it appears to know about you came from the prompt. This is what retrieval-augmented generation exists to fix.
Its knowledge has a cutoff. Training finished on a date. Events after it are invisible unless you supply them.
It is confident either way. The model has no internal signal distinguishing "I learned this reliably" from "this fits the pattern." Fluency is not evidence of accuracy.
What an LLM is good at, and what it is not
| Reliable | Unreliable without help |
|---|---|
| Rewriting and summarizing supplied text | Recalling specific facts about you |
| Extracting structured fields from documents | Arithmetic and multi-step calculation |
| Classifying and routing by intent | Anything time-sensitive or current |
| Translating between formats and vocabularies | Citing sources it was not given |
| Drafting from a clear brief | Consistent output across identical runs |
The pattern in the left column: the information is in front of it and the task is transformation. The pattern on the right: the model is being asked to be a database. Design accordingly. Give it the material, and let a real system handle the arithmetic and the facts.
Choosing a model is the smallest decision you will make
Model choice absorbs a disproportionate share of attention. In practice, for most business tasks the leading models are close enough that the difference is swamped by everything around them: whether the model has your context, how narrowly the task is scoped, and whether outputs are validated.
The factors actually worth weighing are usually not benchmark scores. They are where the data is processed and what the provider retains, which in Quebec is a Law 25 and procurement question. They are cost per unit of work at your real volume, latency where a person is waiting, and whether you can switch providers later without rewriting everything. Treat the model as a replaceable component, because it will be replaced, probably within the year.
How Automathing approaches it
We treat the LLM as one component among several. The prompt, the retrieved context, the tools it can call, the evaluation harness, and the interface a person actually uses all matter more to the outcome than which model sits in the middle. We build so the model can be swapped without redoing the system, because the frontier moves faster than any procurement cycle.
Frequently asked questions
What does LLM stand for?
Large Language Model. "Large" refers to the number of parameters and the volume of training data; "language model" means a system that predicts likely continuations of text. The name describes the mechanism, not the capability.
What is the difference between an LLM and generative AI?
Generative AI is the broader category of models that produce new content, including images, audio, video, and code. An LLM is the language-specialized case. Most business generative AI today is LLM-based, which is why the terms get used interchangeably even though they are not the same.
Can an LLM access our company data?
Not on its own. It has to be connected, through retrieval over your documents, through tools that query your systems, or by placing the relevant records directly into the prompt. This connection work, not the model, is where most of the effort in a real project goes.
Why does the same question get different answers?
LLMs sample from a probability distribution rather than returning a fixed result, so identical inputs can produce different phrasings. You can reduce variation by lowering the sampling temperature, constraining the output format, and narrowing the task, but you cannot make it fully deterministic. If your process requires identical output every time, that step belongs in conventional code.
Do bigger models always perform better?
No. Larger models tend to handle nuance and complex reasoning better, but they cost more and respond more slowly. Plenty of business tasks, including classification, extraction, routing, and straightforward summarization, run perfectly well on smaller, faster, cheaper models. Matching model size to task difficulty is one of the more reliable ways to control cost at volume.