The core concepts
LLM (large language model): the engine behind modern AI assistants; software trained on enormous text collections that predicts language, which at scale looks like reasoning and writing. Generative AI: AI that creates content (text, images, code) rather than only classifying or predicting. Prompt: your instruction to the AI; the quality lever most people underuse. Context window: how much text the AI can consider at once; bigger means whole documents and long conversations. Hallucination: confident, wrong output; the reason human review exists. Token: the unit AI reads and bills in; roughly three-quarters of a word.
The applied layer
RAG (retrieval-augmented generation): letting AI look answers up in your documents before responding; the technique behind “chat with your knowledge base.” Fine-tuning: additionally training a model on your data; rarely needed at SMB scale, usually RAG suffices. Agent: AI that executes multi-step tasks and uses tools, not just answers. Custom assistant: a configured AI with your instructions and documents; the no-code middle ground. API: the connection that lets software talk to AI models; what your developers or tools use behind the scenes. Inference: running the model to get answers; what you pay for per use.
The buying-conversation terms
No-training guarantee: the vendor promise that your data does not train their models; the key clause in business tiers. Data residency: where your data is stored geographically; relevant for compliance. SSO: single sign-on, letting staff log in via your company account; an admin and security basic. Guardrails: the rules and filters around what an AI system may say or do. Open-weight model: a model you can run on infrastructure you control, for maximum data control. Multimodal: handling images, audio and documents, not just text. Benchmark: standardized model tests; useful signal, poor purchasing criterion. AI Act: the EU’s AI law, relevant the moment you sell into Europe.
FAQ
Do I need to understand the technology to benefit?
No more than you need to understand engines to run a delivery fleet. These terms exist so vendors cannot snow you.
What is the most misused term?
“Agent.” Many products labeled agents are assistants with good marketing. Ask what the system does without a human clicking approve.
What should I learn next?
Whatever your first pilot needs. Vocabulary follows use faster than use follows vocabulary.