Large Language Models

Large language models are AI models trained on vast collections of text and other data to understand and generate language. They learn statistical patterns across billions of examples, which is why they can summarize, draft, translate, classify, reason over context, and converse coherently. The same training process is also why they don't guarantee truth: LLMs generate statistically plausible text, which often happens to be accurate — but can be wrong with equal confidence. Grounding them in organization-controlled sources through retrieval is how organizations get reliable outputs rather than plausible ones.

LLMs are the most widely adopted AI capability in enterprise today, which means their failure modes are also the most widely deployed. The knowledge cutoff problem — where models answer questions about current regulations, prices, or internal policies using outdated training data — affects any deployment that relies on the model's memory rather than retrieval. Vendor data handling terms determine whether confidential content submitted in prompts stays private. Human review requirements vary by use case but rarely disappear. None of these are technical edge cases; they're the baseline governance questions for any LLM deployment.