Glossary · 32 terms · updated 1 November 2026

The AI agent glossary, in plain words

An AI agent is software that reads an input (a message, a document, a record), decides which of a short list of actions applies, carries it out through tools connected to your systems, and asks a person when its rules say so. A chatbot only answers; an automation only moves data when a trigger fires; an agent decides. The terms below are the vocabulary around that one idea.

Every term a vendor will use in the first ten minutes, defined in one or two sentences for someone who runs a business rather than builds software. Where the site shows the thing in action, the entry links to it.

The core idea

AI agent
Software that reads a situation, decides which action from a fixed list applies, does it through tools, and escalates to a person when the rules say so. Different from a chatbot (answers only) and from an automation (fixed steps, no reading).Agent vs chatbot vs Zapier
Agentic AI
The general term for systems built around agents: planning, tool use, memory and a loop that keeps going until the task is done or a rule says stop.Agentic AI development
Workflow
One repeated job with a trigger, steps and an outcome: a booking request, a document to file, a report to assemble. Agents are built one workflow at a time.50 workflows
Large language model (LLM)
The model that reads and writes language inside an agent: Claude, GPT, Gemini. It is the engine, not the agent; the agent is the rulebook, the tools and the loop around it.
Prompt
The written instructions the model is given: who it is, what it may do, what it must never do, in what tone. In an agent this is where the rulebook lives.
Rulebook
The written rules an agent applies: which doctors, which slots, how many reschedules, what to say when a message mentions a symptom. The rulebook is the product; the model applies it to messy input.The scoping template

How an agent acts

Tool use (function calling)
The mechanism by which a model asks to run a named function (check availability, book a slot, send a form) with specific inputs, gets the result back, and continues. Tools are how an agent touches your systems.Watch tool calls in the demos
Tool list
The named set of things an agent is allowed to do. It is also the permission list: an agent cannot do what it has no tool for, however it is asked.Tool lists in the case studies
Agent loop
Read the input, decide, call a tool, read the result, decide again, until the task is done or a rule says stop. Capped in turns so it cannot run away.
Autonomy level
How much the agent does alone: draft for a person, act with approval, or act within rules. Builds usually start at the first and earn the third.How autonomy moves the price
Approval queue (human in the loop)
The place where actions that matter wait for a person’s yes before they happen. Every agent on this site starts with one.
Escalation
The agent handing a conversation or a case to a person, with the transcript and a reason, because the rules say this is not its call.
Guardrails
Everything that constrains an agent: the tool list, the approval queue, the escalation rules, the frequency caps, the logging. Guardrails are design, not hope.
Memory
What an agent carries between runs: last cycle’s commitments, a client’s profile, a conversation so far. Usually a database table, not something mystical.
Orchestration
Coordinating several agents or steps: one agent qualifies, another books, a third updates the CRM. Needed later, not first.

How an agent knows things

Retrieval-augmented generation (RAG)
Giving the model the relevant pages from your own documents at the moment it answers, so it replies from your material rather than general knowledge, and can cite where the answer came from.RAG explained
Knowledge base
The approved material an agent may answer from: hours, fees, policies, course modules. Owned by the business, changed without a developer.
Context window
How much text the model can consider at once. Large today, but still the reason an agent is given the relevant documents rather than all of them.
Grounding
Tying every answer to a source the agent was given. An ungrounded agent improvises; a grounded one says "that is not in the material" and escalates.
Hallucination
A confident, wrong answer the model produced because nothing stopped it. The fix is grounding, a narrow tool list and a person approving what matters, not a better model.Hallucination risk
Confidence score
A number an agent attaches to a decision (which ledger a bank line belongs to, which document this is) so that low-confidence cases go to a person.

Channels and plumbing

WhatsApp Business API (Cloud API)
Meta’s official interface for businesses to send and receive WhatsApp messages at scale. Needs a verified business, approved templates and opt-in; priced per message.WhatsApp numbers
Message template
A pre-approved WhatsApp message a business may send to start a conversation (a reminder, a checklist). Free-form replies are allowed inside a window after the customer writes.
API
The doorway through which software talks to other software. A system with an API can be connected directly; one without needs exports or a sheet in the middle.
Webhook
A system telling another system that something happened, the moment it happens: a message arrived, a booking was cancelled. What wakes an agent up.
Integration
Connecting the agent to a system you already run: CRM, calendar, practice software, accounting. Each one is work, which is why the pilot caps it at two.The pilot
Edge function (serverless)
A small piece of code that runs on demand in the cloud with no server to maintain. Where an agent loop usually lives at small-business scale.

Buying and running one

Pilot
A bounded first build: one workflow, two systems, an approval step, a measured result, a fixed fee, two to three weeks.The AI Agent Pilot
Baseline
The number before the agent: hours, response time, no-show rate, completion rate. Measured first, or nobody will know whether it worked.
Running cost
What an agent costs per month after launch: model usage, hosting, messaging fees. Scales with volume; usually tens of dollars at small-business scale.Running-cost bands
Token
The unit model usage is billed in, roughly three-quarters of a word. A dozen-turn conversation with a few tool calls costs a few rupees.
Handover
Prompts, tool code, runbooks, the data-flow diagram and the accounts, delivered so the business runs the agent without the builder.How we work
Data-flow diagram
The two-line picture of what data the agent sees, where it goes and who can read it. The thing to show a customer, an auditor or a regulator who asks.

Cite this page: SigmaticAI, "The AI agent glossary for business owners", appsigmaticai.com/library/ai-agent-glossary.

Link to it rather than copying the table; the numbers are refreshed here and the sources are one click away.