The one-paragraph version
A large language model is a neural network trained on enormous amounts of text to do one thing: predict the next word given everything before it. Done at sufficient scale, that single skill turns out to encompass drafting, summarising, translating, answering questions, extracting structure from messy documents and writing working code. The models behind modern AI assistants are LLMs with additional training to follow instructions and decline harmful requests.
What LLMs are genuinely good at
- Transforming language. Summarise this report, rewrite this for a customer, translate this contract. Reliable, fast, and better than most humans at first drafts.
- Extracting structure. Pulling fields, entities and decisions out of unstructured documents, invoices, emails, contracts, at scale.
- Answering over provided context. Given the right source material, retrieved by a RAG system, LLMs answer grounded questions with citations.
- Code. Generating, explaining and reviewing software under human supervision.
What they are not
An LLM is not a database: it has no reliable record of facts, and it will state falsehoods with the same fluent confidence as truths, the failure mode called hallucination. It is not a calculator: arithmetic and precise logic should be delegated to tools. It does not learn from your conversations by default: without deliberate engineering, it forgets everything between sessions. Each limitation has an engineering answer, retrieval for grounding, tool use for computation, memory systems for continuity, which is precisely why production LLM systems are systems rather than bare API calls.
What running one costs
Commercial LLMs are priced per token, roughly per word processed, which means cost scales with usage rather than with seats. Small tasks route to small, cheap models; complex reasoning routes to large ones. A well-engineered system with routing, caching and prompt discipline routinely runs at a fraction of the cost of a naive one, which is why cost design belongs in the architecture, not in the finance review after launch.
Open models versus APIs
Most companies start with commercial APIs: no infrastructure, immediate access to frontier quality. Self-hosted open-weight models earn their keep when data cannot leave your environment, when per-token costs at high volume exceed hosting costs, or when a model needs deep domain adaptation. Many production systems mix both.
The decision-maker's summary
Treat LLMs as a new computing primitive for language, transformative when wrapped in retrieval, evaluation, guardrails and cost control, unreliable when used bare. The wrapper is where projects succeed or fail. Building that wrapper is our LLM development service in a sentence, and we are glad to show you what it looks like against a concrete use case.