Twelve AI terms every buyer should know before the next vendor meeting

Vendor pitches lean on shorthand. These twelve terms cover most of what you will hear, defined plainly, each with a question worth asking.

1. Large language model (LLM). A model with many parameters, trained on vast amounts of text, that can generate text, answer questions and summarize ^[1]. Ask: which model, and can we switch?

2. Token. The smallest unit a model processes; it can be a word, subword, character or byte ^[1]. Ask: how is pricing counted, per input and output token?

3. Context window. The text a model can reference when generating, its "working memory", distinct from its training data ^[1]. Ask: will our documents fit?

4. Pretraining. The initial training on a large unlabeled corpus, where the model learns to predict the next word ^[1]. Ask: what is in the training data?

5. Fine-tuning. Further training of a pretrained model on additional data to adapt it to a domain, task or style; it needs careful choice of data ^[1]. Ask: is it needed, or is retrieval enough?

6. RLHF. Reinforcement learning from human feedback: people rank outputs, and training steers the model toward preferred ones ^[1]. Ask: how was the model tuned to follow instructions?

7. RAG (retrieval-augmented generation). Relevant documents are retrieved at runtime and passed into the context window so answers are grounded in evidence; quality depends on the knowledge base ^[1]. Ask: what is retrieved, and how is it kept current?

8. Temperature. A setting for output randomness; even at 0, results are not fully deterministic ^[1]. Ask: how do we validate outputs?

9. Latency. The delay between submitting a prompt and receiving output, affected by model size, hardware and prompt complexity ^[1]. Ask: what is typical and worst-case?

10. Time to first token (TTFT). How long until the first output token appears, a key measure for interactive tools ^[1]. Ask: does it meet the user experience we need?

11. MCP (Model Context Protocol). An open protocol that standardizes how applications give context and tools to models, likened to a USB-C port ^[1]. Ask: which tools will it connect to, and who vets the servers?

12. Agent vs. workflow. Workflows orchestrate models and tools through predefined code paths; agents let the model direct its own process and tool use ^[2]. Ask: does this task really need an agent, given the added cost and risk of compounding errors? ^[2]

Practical implication

Bring this list to your next vendor meeting and ask the italicized question for each term the vendor uses. Vague answers on data, validation and who controls tool access are the ones to follow up.