The words you will hear: Start here, ProvisoIQ

Find a term

The basics

Large language model (LLM): The kind of AI behind tools such as Claude, ChatGPT, Copilot and Gemini. It predicts text from patterns learned in training, which is why it can be fluent and still wrong.
Prompt: The instructions and material you give the tool. Better prompts give clearer, more checkable results.
System prompt or instructions: Standing instructions that apply to every message in a chat or project, such as your role, house style and positions.
Hallucination: A confident statement that is not supported by the source or by fact, such as an invented clause or citation.
Training data: The material a model learned from. It has a cutoff date, so the model may not know recent law or events.
Grounding: Tying answers to a specific source, such as the contract you provided, so they can be checked against it.

Context window: How much text the tool can consider at once, including your documents and the conversation so far. Very long material may be cut off.
Token: The unit a model counts text in, roughly three quarters of a word. Limits and prices are often stated in tokens.
Multimodal: Able to work with more than text, such as images, scanned pages or audio.
Model version: The specific model in use. Results can change when the version changes, so note it when you test.
Temperature: A setting that controls how varied the answers are. Lower values give more consistent output.
Retrieval (RAG): A setup where the tool searches a document collection and answers from what it finds, instead of from memory alone.

Training on your data: When a provider uses your inputs to improve its models. Whether this happens depends on the plan and settings.
Retention: How long the provider keeps your conversations and files, and whether you can delete them.
Zero data retention: A setting or agreement under which the provider does not store your inputs after processing, usually on business plans.
Data processing agreement (DPA): A contract setting out how a provider handles personal data on your behalf.
Consumer versus business plan: Personal plans and business or enterprise plans often have different data, training and confidentiality terms.
Acceptable use policy: Your organization's rules on which AI tools may be used, by whom and for which material.

Project or workspace: A saved space that keeps files and instructions together across chats.
Memory: A feature that lets the tool remember facts about you across conversations. Check what it stores.
Plugin or skill: A packaged set of instructions or tools that adds a capability to an AI product.
Tool use or connectors: Links that let the AI read from or act in other systems, such as your email, documents or a database.
Model Context Protocol (MCP): An open standard for connecting AI tools to other systems and data sources.

Agent: An AI setup that takes several steps on its own toward a goal, such as reading files, drafting and checking, usually with points where you approve.
Human in the loop: A person reviews and approves the output at set points before anything is relied on or sent.
Prompt injection: Hidden instructions inside a document or web page that try to make the AI do something you did not ask.

Playbook: Your standard, fallback and walk-away positions for a contract type. Giving one to the AI makes its review match your standard.
Extraction: Pulling specific information out of documents, such as dates, parties or renewal terms, into a table.
Summarization: Condensing a document. Check what was left out as well as what was said.
Citation checking: Confirming that every case, statute or clause the AI cites exists and says what it claims.
Redline generation: AI-suggested changes to a document, shown as tracked changes or a list. Check that each change is the minimum needed.
Benchmark: A standard test used to compare models. Your own tasks are a better guide to your results.

The same 32 terms are on a two-page printable sheet: AI terms for legal work (PDF).