Core concepts
What is LLM?
A large language model (LLM) is an AI model trained on enormous amounts of text to predict language, which in practice lets it read, write, summarise, translate, extract, and reason across almost any text you give it.
GPT-5.2, Claude, and Gemini are LLMs. Under the hood they predict the next token (roughly, the next word fragment) given everything before it. That sounds too simple to be useful, but at sufficient scale it produces systems that can draft a contract clause, read 40 invoices and produce a summary table, or hold a support conversation in Lebanese Arabic. The capability is general; the business value comes from pointing it at one specific job with the right context and constraints.
Two properties matter commercially. First, LLMs are priced per token, so cost scales with how much text goes in and out. Architecture decides whether a feature costs $0.002 or $0.20 per use. In our own product Compumeal, an AI-refined meal plan costs $0.018 to generate because the pipeline sends the model only what it needs and uses smaller models for the cheap steps. Second, LLMs are probabilistic: the same input can produce different output, and they can be confidently wrong. Production systems wrap them in validation, retrieval, and fallbacks rather than trusting raw output.
When a vendor says "we use AI", they almost always mean they call an LLM. The questions that separate serious builders: which models, why those, what happens when the model is wrong, and what does a unit of work cost at your volume? Anyone who cannot answer those four has not shipped much.
Frequently asked
- What is the difference between an LLM and generative AI?
- Generative AI is the broad category of models that produce new content, including images, audio, and video. An LLM is the text-and-language member of that family: GPT-5.2, Claude, and Gemini. When a business vendor says "we use AI", they almost always mean they call an LLM.
- How much does it cost to use an LLM?
- LLMs are priced per token, so cost scales with how much text goes in and out. Architecture, not the model alone, decides whether a feature costs $0.002 or $0.20 per use. Our Compumeal meal-plan generation runs at $0.018 because the pipeline sends the model only what it needs and uses smaller models for the cheap steps.
- Can LLMs be trusted to be accurate?
- Not on their own. LLMs are probabilistic and can be confidently wrong, so raw output is not something to build on directly. Production systems wrap them in retrieval from your real data, validation against your database, and fallbacks for when the model is unsure. The trust comes from the system around the model, not the model by itself.
Wondering where this fits in your business?
The free AI audit maps your operations and shows the three systems we would build first, in plain language like this.
