AI Terms for Business Leaders: A Plain-Language Glossary

An AI glossary for business is a short list of the terms that appear in proposals, defined by what they change for your company rather than how they work inside. The ones that matter most are agent, harness, context, retrieval-augmented generation (RAG) and tool call, because each points to a decision you will be asked to make: what the system may do, what it may know, and who can watch and stop it. This article defines them in plain language and gives the question to ask behind each one.

Key takeaways

• Most AI jargon in a proposal reduces to three questions: what can the system do, what does it know, and who can see and stop it.
• An agent is a model given a goal and tools. A harness is the software around the model that runs it, limits it and records what it does.
• Context is what the model sees at the moment it answers. RAG is one way to place the right company documents into that context.
• A tool call is the point where AI touches your real systems, so it is where permissions and approvals belong.
• If a term cannot be explained in one sentence and tied to a test, treat it as a gap to close before you sign.

What do the basic building blocks mean?

Before the agent vocabulary, a handful of foundation terms come up in nearly every conversation. You do not need the mathematics behind them. You need to know which part of your business each one affects.

• Model: the trained system that reads input and produces a response. Models differ in quality, speed and running cost, and a company can use more than one.
• Large language model (LLM): a model trained on large amounts of text to understand and generate language. Most business AI products are built on one.
• Prompt: the instructions and material given to the model for a single request.
• System prompt: the standing instructions that define an assistant's role, tone and rules. Think of it as a written job description the model reads every time.
• Token: the small piece of text a model reads or writes. Usage is often measured in tokens, which is how the term connects to your invoice.
• Context window: the most text a model can consider at one time. Anything beyond it must be trimmed, summarized or left out.
• Hallucination: a confident answer that is not supported by facts or sources. Good retrieval, clear rules and review reduce it. Nothing removes it entirely.

A useful habit is to ask whether a term describes the model or the system around it. The model is often replaceable. The system around it, with its memory, tools and rules, is what you will live with, so most of your attention belongs there.

What are an agent, a harness and a tool call?

These three terms describe how AI moves from answering questions to doing work. For a fuller treatment of the first, see what an AI agent for business is.

• Agent: a model given a goal, instructions and a set of tools, which decides its next step, takes it, checks the result and continues until the task is done or it needs help. A chatbot answers. An agent acts within limits you set.
• Workflow: a fixed sequence of steps, some of which may use a model. Predictable and easy to test. Many useful systems are workflows with one or two flexible steps, and that is often the right design.
• Harness: the software around the model that runs the loop. It assembles what the model sees, executes the tools the model asks for, enforces limits, handles errors and records what happened. When a proposal says the agent is safe, the harness is usually where that claim is either true or empty.
• Tool: a defined action the agent is allowed to use, such as searching a knowledge base, reading a calendar or drafting a reply in a helpdesk.
• Tool call: the moment the model asks for a tool to be used, with specific inputs. The harness decides whether to run it, runs it and returns the result. The model never touches your systems directly.
• Model Context Protocol (MCP): an open standard for connecting models and agents to tools and data sources in a consistent way. It reduces custom wiring but does not decide what the agent is permitted to do.
• Orchestration: coordination of several steps, tools or agents toward one outcome.

The practical consequence is simple. Every tool call is a place where something can go right or wrong in the real world: a message sent, a record changed, a file read. So the questions that matter are which tools exist, what each one can touch, and which calls need a person to approve them first.

How do context, memory and RAG fit together?

A model knows only what it learned in training and what is placed in front of it right now. Most of the work in a business agent is deciding what to place in front of it.

• Context: everything the model sees for a given step, including instructions, the conversation, retrieved documents and tool results. Because the context window is limited, what goes in is a design choice.
• Context engineering: the practice of selecting, ordering and trimming what enters the context so the model has what it needs and little else. More text is not better. Irrelevant material can make answers worse.
• Memory: information kept between sessions so the agent does not start from zero, such as preferences, past decisions or the state of a long task. Ask where it is stored, who can read it, and how it is corrected or deleted.
• Retrieval-augmented generation (RAG): a method where the system first searches your documents for relevant passages, then gives them to the model to answer from. The answer can point back to a source, and updating a document updates the answer.
• Embeddings and vector search: a way of representing text so that passages with similar meaning can be found even when the words differ. This is usually the search engine inside a RAG setup.
• Fine-tuning: further training of a model on examples to change how it behaves or writes. It teaches style and patterns more than it keeps facts current.

For company knowledge that changes, such as policies, price lists of products, procedures and customer histories, retrieval is usually the first approach to evaluate, because the documents can be updated without retraining anything. Fine-tuning is more often considered for consistent format or tone. Neither fixes poor source documents. If the policy folder contradicts itself, the agent will too.

Where these pieces sit in a complete system is easier to see on a map. The seven layers of company AI infrastructure place memory, tools and controls in one picture.

Which terms describe control and safety?

This group deserves the most attention from owners, because it is where business risk lives and where vague language is most common.

• Permissions: what the agent's account is allowed to read and change. The sound principle is least privilege: only what the role needs, nothing extra.
• Approval boundary: the line between actions the agent may take alone and actions that wait for a named person. Drafting an email may be inside the line. Sending a payment is usually outside it.
• Human in the loop: a person reviews or approves before an action takes effect. It is a design choice you can apply to some steps and not others.
• Guardrails: rules and checks that limit what the agent can say or do, such as blocked actions, required formats or topic limits. Ask whether they are enforced by the harness or only requested in the prompt, since the second kind can be talked around.
• Prompt injection: an attack or accident where text the agent reads, such as an email or web page, contains instructions that try to redirect it. Defenses include limiting tools, separating trusted instructions from untrusted content and requiring approval for sensitive actions.
• Audit trail: a record of what the agent saw, decided and did, kept so a person can reconstruct events later.
• Data residency: where company data is stored and processed, which can matter for contracts and regulation.

When a proposal says secure or compliant, translate it into these words. Which permissions does the agent hold? Which actions need approval? What is logged? A permission review is the routine that keeps those answers true after launch, not only on day one.

Which terms describe measurement and operations?

A system you cannot measure is a system you can only trust or distrust. These terms turn that into something you can check.

• Evaluation (eval): a repeatable test of how the agent performs on realistic tasks, with a defined way to judge the result. Evals are what let you change something and learn whether it got better or worse.
• Success criterion: the observable outcome that says the agent is doing its job, agreed before launch. See how to define success criteria for an AI agent pilot.
• Observability: the ability to see what the system did and why, through logs and traces of each step, tool call and result.
• Latency: how long a response takes. It matters most where a customer or employee is waiting.
• Usage limits: caps and alerts on consumption so a loop or a busy week does not produce a surprise.
• Regression: something that used to work and stopped after a change, such as a new model version or an edited instruction.
• Drift: gradual change in results because the world moved: documents age, products change, people ask different things.

Notice that none of these are features you buy once. They are habits of operation. A proposal that describes the build in detail and operations in one line has left out the part that decides whether the agent is still good in a few months.

How do you use this glossary when a proposal arrives?

A glossary only helps if it changes how you read a document. Here is a short method you can apply to any proposal or demo.

• Underline every term from this article that appears in the proposal.
• Next to each, write the one-sentence definition in your own words. If you cannot, ask the vendor to do it in writing.
• For each tool, ask what it can read, what it can change and which calls need approval.
• For each source of knowledge, ask who maintains the documents and how an outdated one is corrected.
• For each claim of quality, ask which evaluation supports it and whether you can see it run on your own examples.
• For each ongoing promise, ask who does the work, how often, and what you receive as evidence.

Illustrative scenario: a proposal for a services firm describes an autonomous agent with RAG, MCP tools and guardrails. Reading it with the glossary, the owner asks four things. Which documents does retrieval search, and who keeps them current? Which tools can write to the client database, and are those calls approved by a person? Are the guardrails enforced in the harness or only written in the prompt? Which evaluation will be run before launch, and again after each change? None of the questions is technical. Together they show whether the vendor has designed a system or assembled a demo.

How does Be Human handle these terms in practice?

Be Human designs, implements and operates personalized AI infrastructure for companies, so each of these terms corresponds to a defined component rather than a slogan. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, while identity, memory, channels, tools and permissions stay under human control. Service is remote, in English and Portuguese.

Frequently asked questions

What is the difference between an AI agent and a chatbot? A chatbot responds to messages, usually from a fixed body of knowledge. An agent has a goal, tools and the ability to take steps in your systems within permissions you define, which is why it needs approval boundaries and logging that a simple chatbot may not.

Do I need RAG for my company's AI? You need it if the agent must answer from your own documents, especially ones that change. If the task depends only on general knowledge or on live data from a tool, retrieval may be unnecessary. Decide by listing what the role must know and where that information lives today.

What should I ask about a harness? Ask what it enforces on its own, such as limits, approvals and logging, as opposed to what it merely asks the model to do. Also ask whether you can see its records and whether it works with more than one model, since that affects how freely you can change providers later.

How much of this vocabulary do I need to learn? Enough to ask for plain-language answers and recognize when you do not get them. If you can explain what the agent can do, what it knows and how its work is checked, you know enough to decide. The details can stay with the people operating the system.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.