Tier one: the baseline
If you do not know these, you cannot evaluate anything. Start here.
AI model
The engine underneath every AI product. A model is trained on large amounts of data and learns to recognize patterns, generate text, make predictions, or take actions. When a vendor says "our AI," they mean a model doing something specific.
What it means to you: the model is not the product. The product is what someone has built on top of it. Ask what the model is actually being used to do, and how it is being measured.
Large language model (LLM)
The category of AI model behind tools like ChatGPT, Claude, and Gemini. Trained on vast amounts of text, they can read, write, summarize, analyze, and reason across almost any topic.
What it means to you: this is the technology your teams are already using informally, whether you know it or not. The question is whether you have a strategy around it or just shadow usage with no governance.
Chatbot
A conversational interface. Something you type into and get a response from. Chatbots can be simple (scripted, rule-based) or sophisticated (powered by an LLM). Most of what people call "AI" in the consumer world is a chatbot.
What it means to you: a chatbot is an interface, not a strategy. It can be useful for customer service, internal Q&A, or onboarding. But a chatbot that cannot take action or connect to your systems is just a fancy search box.
AI agent
An AI that does not just respond. It acts. An agent can browse the web, run a search, write and execute code, send an email, update a database, or trigger a workflow. All on its own or in sequence. Agents are what make AI genuinely operational.
What it means to you: this is where business value accelerates. An agent can handle a multi-step process, pulling data, analyzing it, drafting a report, and sending it to the right person, without a human touching it at every stage. When someone talks about AI automation, they are usually talking about agents.
Prompt
The instruction you give an AI. Every output from an AI model starts with a prompt. A question, a command, or a set of instructions that tells the model what to do.
What it means to you: the quality of your prompts determines the quality of your outputs. This is why "just use ChatGPT" is not a strategy. How your teams are prompting AI models, and whether they are doing it consistently and well, directly affects the value you get.
Hallucination
When an AI generates something that sounds confident and coherent but is factually wrong. It does not know it is wrong. It produces the most plausible-sounding answer, not necessarily the correct one.
What it means to you: every AI output that touches a customer, a contract, a financial model, or a regulatory requirement needs a human review step. Hallucination is not a bug that will be fixed. It is a property of how these models work. Build your processes accordingly.
Training data
The information an AI model learned from. A model trained on general internet data knows a lot about the world but nothing about your business. A model trained on your data knows your business.
What it means to you: the most durable competitive advantage in AI is proprietary data. Companies that figure out how to use their own operational data, customer behavior, production records, sales history, research outputs, will build systems their competitors cannot replicate.
Tier two: what you need to evaluate a proposal
The concepts that separate a real implementation from a demo that will never ship.
Retrieval-augmented generation (RAG)
A technique that connects an AI model to a specific knowledge base. Your documents, your data, your internal systems. So it can answer questions using your information rather than its general training.
What it means to you: this is how you build an AI that knows about your business without rebuilding the model from scratch. An AI that can search your product documentation, your client history, or your operational data and answer questions from it. When a vendor says their AI is "connected to your data," RAG is usually how.
Questions to ask: what data is it connected to, how current is that data, and how do you control what it can and cannot access?
Fine-tuning
Taking an existing AI model and training it further on your specific data so it behaves in a way that is more relevant to your context. Your industry language, your processes, your outputs.
What it means to you: fine-tuning makes a general model more specialized. It is useful when you need the AI to consistently match a tone, follow specific rules, or perform a narrow task at high quality. It is also more expensive and time-consuming than RAG. Most mid-market businesses do not need it to start.
Questions to ask: are you recommending fine-tuning because it is the right solution, or because it is what you know how to sell?
Automation vs augmentation
Two distinct ways to deploy AI. Automation replaces a human step entirely. The AI does the task without a person involved. Augmentation gives a person better information or tools to do their job faster and with higher quality.
What it means to you: automation has a ceiling. You can only remove so many steps before you hit compliance, relationship, or judgment constraints. Augmentation compounds. Every person operating with better information and less friction produces better outcomes over time. Most businesses need both, but the balance matters.
Questions to ask: is this designed to remove people from the process or to make people in the process better? And which one does our situation actually call for?
API (application programming interface)
The connection point between two software systems. APIs let AI tools plug into your existing platforms. Your CRM, your ERP, your data warehouse. So they can read information and take action inside systems you already use.
What it means to you: an AI that cannot connect to your systems via API is an island. It can produce outputs but cannot act on them. Before committing to any AI implementation, map which of your core systems have APIs and which do not. That is your constraint map.
Questions to ask: what systems does this connect to out of the box, and what does custom integration cost and take?
Context window
The amount of information an AI can hold in its working memory during a single conversation or task. A small context window means the AI forgets earlier parts of a long document or conversation. A large one lets it reason across much more at once.
What it means to you: context window size matters when you are asking AI to analyze long documents, complex datasets, or extended workflows. If a vendor's AI cannot handle the length of your actual contracts or reports, that is a practical constraint worth testing before you commit.
Questions to ask: what is the context limit, and have you tested it against our actual documents?
Inference
The moment when an AI model is actually running. Taking an input and producing an output. Training happens once (or periodically). Inference happens every time someone uses the model.
What it means to you: inference costs money and takes time. High-volume use cases, processing thousands of documents, running AI on every customer interaction, need a cost and latency model built in from the start. Ask what inference costs at scale before you design around it.
Questions to ask: what does this cost per query at our expected volume, and how does that change as we scale?
Workflow integration
The degree to which an AI tool is embedded into the way your team actually works, rather than sitting alongside it as a separate application people have to remember to use.
What it means to you: adoption lives or dies here. An AI tool that requires people to change their habits, open a new application, or manually move information in and out will not get used. The highest-value implementations are invisible. AI working inside the tools and processes your team already relies on.
Questions to ask: does this live inside our existing workflow or does it require a behavior change to access it?
Tier three: worth knowing, but let your team own the depth
You do not need to go deep here. You need enough to know when someone is overcomplicating things.
Tokens
The units AI models use to process text. Roughly three-quarters of a word each. Models have token limits per request and charge by token usage.
What it means to you: tokens are the unit of cost and capacity. Know that they exist. Let your team manage the specifics.
Embeddings
A way of representing information, text, images, data, as numbers so an AI can measure similarity and relevance between pieces of content. The foundation of search and recommendation systems.
What it means to you: if someone is building a system that needs to find relevant information from a large body of content, embeddings are how it works. You do not need to understand the math. You need to know it is a solved, reliable technique.
Vector database
A type of database optimized for storing and searching embeddings. Used in RAG systems to let AI retrieve the right information quickly.
What it means to you: this is infrastructure. If your team is building a serious AI system, they will need one. It is a commodity choice. Do not let a vendor charge you a premium for it.
Parameters / model size
A measure of an AI model's complexity and capability. Larger models generally perform better on complex tasks but cost more to run. Smaller models can be faster and cheaper for narrow, well-defined tasks.
What it means to you: bigger is not always better. A smaller, faster model that does one thing well may outperform a general large model for a specific use case, and cost a fraction of the price. Ask whether the model is sized right for the task, not just whether it is the most powerful option available.
Frontier labs
The small group of companies doing the most advanced AI research and building the most capable models: Anthropic, OpenAI, Google DeepMind, and Meta AI. They set the pace for what is possible and release new model versions every few months.
What it means to you: the frontier moves fast. A capability that did not exist six months ago may be standard today. Any vendor or internal team working with AI should be tracking what the frontier labs are releasing. And you should be asking how their roadmap changes what is available to your business.
Open source models
AI models whose underlying code and weights are publicly available. Anyone can download, run, and modify them. Leading examples include Meta's Llama series and Mistral. Distinct from proprietary models (OpenAI, Anthropic, Google) which are accessed via API and controlled by the vendor.
What it means to you: open source gives you more control over your data and lower long-term costs. Nothing leaves your infrastructure. The tradeoff is that open source models are generally less capable than the frontier proprietary ones, and running them requires more technical investment. For data-sensitive industries like biotech or financial services, open source is worth serious consideration.
Local LLM
An AI model that runs entirely on your own hardware. A server you control, on your premises or in your private cloud, rather than sending data to an external vendor's servers.
What it means to you: if data privacy, regulatory compliance, or competitive sensitivity is a concern, local LLMs let you use AI without your information ever leaving your environment. The capability gap versus cloud-based frontier models is narrowing. This is not a fringe option anymore. It is a legitimate architecture choice for businesses with strict data requirements.
Model context protocol (MCP)
A standard that lets AI models connect to external tools, data sources, and services in a consistent way. Think of it as a universal plug for AI. Instead of building a custom connection between an AI and each system it needs to access, MCP provides a common language that works across tools.
What it means to you: MCP is what makes AI agents practical at scale. Without a standard like this, connecting an AI to your CRM, your ERP, your databases, and your communication tools requires custom engineering work for each one. With it, the connection is faster, more reliable, and easier to maintain. If a vendor is building you an agent-based system, ask whether they are building on MCP or proprietary integrations, and why.
Part of the Proof of Value series from Busted Eye. Let us know if you have questions about what any of this means for your business and how we can help.