Plain-English definitions of the AI, automation, and operations terms every business operator should actually know. No jargon, no hype, no marketing language.
An AI system that operates autonomously toward a goal, taking multiple steps without human input in between. Unlike an AI assistant, an agent can call tools, make decisions, and continue until the goal is met or a stopping condition is hit.
Most products marketed as "AI agents" today are actually assistants in marketing copy. The honest check: can it complete a real workflow end-to-end with no human watching?
An AI system that responds to a prompt and then waits for the next one. Conversational by nature, with a human in every loop. Examples include ChatGPT, Claude.ai, and Perplexity.
The opposite of an AI agent. Assistants are reactive; agents are proactive.
An AI assistant built by OpenAI, accessible through chat.openai.com or the OpenAI API. Powered by the GPT series of large language models. Strong on real-time web search, image generation, voice mode, and broad consumer reach.
The default consumer AI for most teams. Often paired with Claude for different parts of the workflow.
An AI assistant and API built by Anthropic. The Sonnet, Opus, and Haiku models can be accessed through claude.ai (consumer chat) or the Anthropic API (developer access). Strong on long-context reasoning, tone consistency, and structured outputs.
Our default LLM for production workflows at Innovative Compass. Paired with ChatGPT on most client builds.
A numerical representation of text (or other data) that captures its meaning. Words and phrases with similar meanings have similar embeddings. The math foundation of semantic search and RAG.
Embeddings are what get stored inside a vector database, then queried when an AI needs to find relevant context.
Taking a pre-trained AI model and continuing training on a smaller, specialized dataset to make it better at a specific task.
Less common in 2026 than it used to be. Most teams now use prompt engineering or RAG instead because they're cheaper, faster to iterate, and avoid the complexity of managing a custom model.
The practice of optimizing a website's content and structure so it gets cited by AI search engines like ChatGPT, Claude, and Perplexity. The successor concept to SEO.
Focuses on scannable summaries (TL;DR blocks), structured data (JSON-LD), citation-friendly content patterns, and explicit definitional sentences. Different from SEO because the goal isn't ranking in a list of links, it's being quoted directly in an AI's answer.
A type of AI model trained on massive text datasets to understand and generate human language. Powers products like Claude, ChatGPT, and Gemini.
Called a "language model" because the math operates on sequences of words (or tokens) rather than images, audio, or structured data.
An open standard developed by Anthropic for connecting AI assistants to external tools and data sources. Lets a single LLM access multiple systems through one consistent interface.
Often described as the "USB-C for AI." Solves the integration problem between LLMs and the tools they need to act on in the world.
An open-source workflow automation platform. Self-hostable, code-friendly, and free if you run it yourself.
Our default choice for production-grade automation builds where you want full control of data and cost structure. Compared to Zapier and Make in our three-way breakdown.
The practice of writing inputs to an LLM that consistently produce the desired output. Includes techniques like few-shot examples, chain-of-thought reasoning, system messages, and structured output formats.
Less of a formal discipline now than a skill operators need. The teams getting the most from AI in 2026 aren't the ones with the cleverest prompts, they're the ones who've turned good prompts into repeatable system prompts.
A pattern where an LLM is given relevant documents (retrieved from a database) before answering a question, so its responses are grounded in specific source material rather than general training data.
Used heavily for building AI assistants on proprietary content: internal wikis, customer support knowledge bases, product documentation. Requires a vector database and an embedding model.
A diagnostic engagement that maps how a business actually operates: what tools it uses, where work stalls, what processes are documented versus invisible. Surfaces inefficiencies before any recommendations get made.
Our standard first step for new clients. We don't recommend a single tool until we've seen the system. Covered in our Strategy Consulting engagement.
The basic unit of text an LLM processes. Roughly 1 token equals 0.75 words in English.
Important because LLM pricing and context limits are measured in tokens, not characters. A 200K-token context window holds roughly 150,000 words. A million-token call at 2 cents per thousand tokens costs $20.
A database designed to store and search embeddings (numerical representations of text, images, or other data).
Used heavily in RAG systems to find documents semantically similar to a query. Examples: Pinecone, Weaviate, pgvector (the Postgres extension).
Software that triggers actions across multiple tools automatically when certain conditions are met. The category includes Zapier, Make, n8n, and custom-built systems.
The point is to remove humans from repetitive operational work so they can focus on tasks that actually require judgment. The core of what we build inside Automation Builds engagements.
A 30-minute call. We'll map these concepts to your actual business and tell you which ones matter for your team right now.