Large language model (LLM)
A large language model, or LLM, is a model trained on enormous amounts of text to read, write, and reason about language. It is the engine behind ChatGPT, Claude, Gemini, and most of the AI features that have shown up in the tools you already use over the past two years. In a business context, an LLM is the part of an AI workflow that handles unstructured language — pulling structure out of free text, drafting responses, summarizing long documents, classifying messages, and answering questions when given the right context.
How LLMs apply in practice
An LLM is not magic and it is not a database. It is a pattern-prediction system that takes input text and produces output text, one piece at a time. That makes it powerful for some tasks and weak for others. Good applied-AI work matches the tool to the task instead of assuming an LLM is the right answer everywhere.
- Summarize and draft. Turn a long email thread into the three things you need to decide. Draft a follow-up. Generate a meeting summary.
- Classify and extract. Read an inbound message, identify what it is about, pull out who/what/budget/timeline.
- Read documents. Pull renewal dates, parties, amounts, and clauses out of contracts, leases, and statements.
- Answer questions in context. When given a set of documents (via RAG), answer questions grounded in those documents instead of guessing.
- Generate structured output. Produce JSON, CSV, or other machine-readable formats from unstructured input.
- Power conversational interfaces. The thing behind every chat sidebar in modern SaaS.
Why LLMs matter
The reason LLMs changed the practical landscape of business software is not the math — neural networks have existed for decades. It is that LLMs became capable enough, fast enough, and cheap enough to run real operational work, not just chat in a sandbox. Tasks that used to require a human to read every input — sorting inbound emails, reviewing invoices, summarizing meeting recordings, extracting data from photos of receipts — can now be done well by software, with humans reviewing the edge cases.
What is still true: LLMs hallucinate, do not always tell the truth, and work best when a human is in the loop for decisions that matter. The companies getting value from them are not deploying them everywhere — they are deploying them carefully, in the places where the model actually changes the answer.
Closely related concepts
Applied AI
What you do with an LLM in a real business.
Agentic workflow
Software that uses an LLM to perform tasks.
Retrieval-augmented generation (RAG)
How LLMs answer questions grounded in your data.
Embedding
The numeric representation of text that LLMs and search rely on.
Fine-tuning
How an LLM gets specialized to your domain.
Document intelligence
A major LLM use case in SMB workflows.
Common questions about LLMs
What can an LLM do well?
Summarizing, drafting, classifying, extracting information, translating, answering questions grounded in provided text, and generating structured output from unstructured input.
What can an LLM do poorly?
Math, exact citation, very long-horizon reasoning, and any task requiring knowledge it does not have. LLMs also confidently hallucinate, which is why grounding and verification matter.
Do you need to fine-tune an LLM to use it?
Usually no. Most applied-AI use cases work with a general-purpose LLM plus prompting and RAG. Fine-tuning is for cases where vocabulary or behavior needs to be locked in.
How does AMG use LLMs?
As one tool among many — sometimes the answer, sometimes overkill. Where it is the right tool, for document extraction, transaction categorization with context, drafting, summarization, and meaning-aware search.
Want to know if an LLM fits your workflow?
See how AMG applies LLMs to operational work — and when we use a smaller tool instead.