Glossary

AI agent and automation glossary: 30 terms, defined

The vocabulary used in every Black Ops scoping conversation, written for the person who has to approve the project.

Updated:

An AI agent reads context, decides, and acts inside your systems. A chatbot only answers. This glossary defines the 30 terms that come up when scoping automation work — agents, RAG, RPA, guardrails, tokens, orchestration, human-in-the-loop — in plain operator language, with no research jargon and no vendor framing.

Which terms does this glossary define?

Thirty, in alphabetical order. Every heading is linkable, so you can send a colleague straight to one definition.

A

Agentic workflow

A process where the sequence of steps is decided at run time by a model instead of being fixed in advance by a developer. The agent reads the current state, picks the next action, executes it and re-evaluates. Classic automation runs the same path every time; an agentic workflow adapts the path to the case in front of it.

AI agent

Software that pursues a defined goal by reading context, choosing an action, calling a tool or system, and checking the result before continuing. The distinguishing feature is action, not conversation: an agent updates a CRM record, posts a journal entry or sends a follow-up. Black Ops builds these as custom agents shaped around one process, and the full definition sits in our guide to what an AI agent for business actually is.

B

Business process automation (BPA)

The practice of running an end-to-end business process — order intake, invoice approval, employee onboarding — with software instead of manual steps. BPA is the outcome; agents, RPA and no-code tools are the means. A BPA project succeeds or fails on process definition and data quality long before the technology matters.

C

Chatbot

An interface that answers questions in a conversation and stops there. It retrieves information and generates text; it does not change the state of your systems. A chatbot that can also update a record, trigger a refund or schedule a delivery has stopped being a chatbot and become an agent. The comparison between agents and chatbots sets out the practical consequences.

Context window

The maximum amount of text a model can consider at once, counted in tokens and covering the system prompt, the conversation, retrieved documents and the model output. When the limit is reached, older content falls out. Window size sets a hard ceiling on how much of a long document an agent can reason over in a single pass.

CRM automation

Software that maintains and acts on the data inside a CRM without a rep doing it by hand: enriching contacts, updating deal stages, logging activity, routing leads, drafting follow-ups. Rules-based CRM automation fires on triggers you define. Agent-based CRM automation reads the deal context and decides what the next action should be.

Custom agent

An agent built around one company’s specific process rather than configured from a template. The decision logic, the integrations and the outputs come from that company’s own rules. Black Ops delivers these with the specification approved before any build starts and deployment in 4 weeks.

D

Deployment

The point at which an agent runs against live data and real work depends on it. Deployment covers credentials, permissions, monitoring, escalation paths and a rollback plan, not just the model. A Black Ops deployment runs 4 weeks from approved scope, with staging review before anything touches production.

E

Embedding

A numeric representation of text that places similar meanings close together in a mathematical space. Embeddings let a system find the three most relevant paragraphs in ten thousand documents without keyword matching. They are the retrieval layer underneath RAG and the reason a vector database exists.

F

Fine-tuning

Training an existing model further on your own examples so it adopts a specific format, tone or classification behaviour. Fine-tuning changes how the model responds; it does not give the model access to current data — that is what retrieval does. Most business automation needs better prompts, tools and retrieval, not fine-tuning.

Function calling

The mechanism by which a model requests that a specific piece of code be run with specific arguments, rather than writing a paragraph about it. The application executes the call and returns the result. Function calling is what turns text generation into a system that can read and write real records.

G

Guardrails

The constraints that define what an agent may and may not do: which systems it can write to, which amounts require approval, which requests it refuses, what it can never send without review. Guardrails are configuration and code, not a paragraph in a prompt. An agent without them is a demonstration, not a production system.

H

Hallucination

Output that is fluent, confident and wrong: an invented invoice number, a policy that does not exist, a citation to nothing. It happens because a language model generates plausible text rather than looking facts up. Retrieval, structured output, tool use and human review reduce it. Nothing removes it entirely.

Human-in-the-loop

A design where a person approves, corrects or overrides specific agent actions before they take effect. The loop is targeted, not universal: the agent handles the routine volume and routes exceptions, high-value decisions and anything below a confidence threshold to a named person. It is the standard configuration for a first deployment.

I

Integration

The connection through which an agent reads from and writes to an existing system — CRM, ERP, ticketing, email, data warehouse. Integrations come in three shapes: a documented API, direct database access, or a scheduled file export. The shape determines whether the agent runs in near real time or on a batch cycle.

L

Large language model (LLM)

A model trained on large volumes of text to predict what comes next, which in practice means it can summarise, classify, extract, translate and draft. An LLM is a component, not a product. On its own it has no memory of your business, no access to your systems and no ability to act.

M

MCP (Model Context Protocol)

An open standard for connecting AI applications to external tools and data sources through a common interface, so a capability built once can be reused by any compatible client. MCP matters operationally because it reduces the number of one-off connectors a deployment has to build and then maintain for years.

Multi-agent system

An arrangement where several specialised agents each handle part of a task and pass work between them: one extracts, one validates, one writes to the system of record. It buys clearer boundaries and easier debugging at the cost of more moving parts. Use it when one agent’s instructions have grown too broad to test.

O

Orchestration

The layer that decides which agent, model or tool runs when, handles retries and timeouts, keeps state between steps and escalates when something fails. Orchestration is the unglamorous part that determines whether an automation survives a bad week. Most failures in production are orchestration failures, not model failures.

P

Prompt engineering

Writing and structuring the instructions given to a model so the output is consistent enough to depend on: task definition, format, examples, constraints and refusal conditions. In production it is a versioned artefact tested against real cases, not a clever sentence. It complements guardrails and retrieval; it does not replace them.

R

RAG (retrieval-augmented generation)

A pattern where the system searches your own documents and records first, then hands the retrieved passages to the model as context for its answer. RAG keeps answers grounded in current internal data and makes them traceable to a source. It is the standard way to make a model useful over a private knowledge base.

RPA (robotic process automation)

Software that repeats a fixed sequence of user-interface actions: open the screen, copy the field, paste it, click submit. RPA is fast, cheap per run and highly predictable on stable systems with structured data. It breaks when the screen, the format or the rule changes, which is also its main maintenance cost. Compared side by side in AI agents vs RPA vs no-code automation.

S

Scope

The written definition of what will be built, what it connects to, what it deliberately will not do, how success is measured and what the investment range is. At Black Ops the scope is written after the first conversation and arrives within 48 hours. Build begins after approval, not before.

Structured output

Model output constrained to a defined shape — a JSON object with required fields and types — instead of free prose. Structured output is what lets a downstream system consume the result without parsing guesswork, and it removes a large class of failures at the boundary between the model and your database.

System prompt

The standing instruction set that defines an agent’s role, tone, permitted actions and refusal rules, applied to every request before any user input. It is where the operating policy of an agent lives. It is not a security boundary: anything that must be enforced belongs in guardrails and permissions.

T

Token

The unit a model reads and writes, roughly a short word or a fragment of one. Tokens are how context windows are measured and how inference is billed, on both input and output. Practical consequence: sending an entire document on every request costs more than retrieving the three paragraphs that matter.

Tool use

An agent’s ability to call external systems — query a database, send an email, create a ticket, run a calculation — and use the result in its next decision. Tool use is the line between a model that describes an action and an agent that performs it, and it is what makes permissions design necessary.

V

Vector database

A store built to hold embeddings and return the closest matches to a query in milliseconds. It is the retrieval half of RAG. For small document sets a vector database is often unnecessary; for hundreds of thousands of records it is the difference between an answer in one second and an answer in thirty.

W

Workflow

The ordered set of steps, decisions, owners and handoffs that carry a process from trigger to completion. Automating anything starts with writing the workflow down as it actually runs, exceptions included — not as the procedure document describes it. Undocumented steps are the most common reason a deployment stalls in week three.

Which of these terms decide the shape of a deployment?

Four of them, in practice. Scope decides what gets built. Integration decides whether the agent runs in real time or in batches. Guardrails and human-in-the-loop decide what it is allowed to do without a person. Everything else on this page is implementation detail underneath those four decisions.

If you want to see the vocabulary applied to a real budgeting exercise, read what drives the cost of automating a process with AI. If you want it applied to a technology choice, read the comparison of agents, RPA and no-code. If you want it applied to a specific region, read how AI automation works across Latin America. The Black Ops FAQ answers the commercial questions these definitions raise.

Next step

Let's talk about
your company

A 30-minute conversation about your company, your process and what you sell. In under 48 hours you receive the brief in writing.

sales@blackopss.com