AI agent glossary

This glossary answers the AI words owners run into when they try to automate the shop. Each page starts with a direct answer, then shows how you would use the idea, a made-up shop example, and when to leave it alone. Synonyms are explained on the main term. The examples are illustrations, not case studies.

A to Z

A
B
C
D
E
I
J
K
L
M
O
P
R
S
T
U
W
Z

Ideas

What the software is doing, before you pick a product or a setting.

AI agent
An AI agent is software that can take a goal, pick the next step, use tools you allow, and keep going until it finishes or it needs you. A chatbot answers. An agent is allowed to do something.
AI receptionist
An AI receptionist is an agent that answers calls or messages, captures the reason, and either books, texts back, or hands the conversation to a person.
Automation vs agent
Automation follows a path you spelled out. An AI agent can choose the next step when the input is messy. Use automation for the same steps every time. Use an agent when the words change and a person would have to judge them.
Chatbot
A chatbot is a conversation window that answers from a script, a knowledge base, or a model. It usually responds when someone types. It does not, by itself, go off and complete a multi-step job.
Computer use
Computer use means an agent operates software through the screen: it can open pages, click, and type, the way a person would. A browser agent is the same idea limited to the web. It is not the same as calling an API.
Context window
A context window is how much text the model can see at once: your prompt, the files you attached, the earlier messages, and the answer it is writing.
Copilot
A copilot is an assistant that works next to you inside a tool, drafting or suggesting while you stay in charge of the file. It is a pattern. Microsoft Copilot is a product family that uses the word. Other products use it too. The pattern is not a promise about any one vendor's features.
Embeddings
Embeddings are lists of numbers that represent the meaning of a chunk of text. Similar meanings land near each other, so software can search by idea instead of exact words.
Fine-tuning
Fine-tuning is additional training that changes a model's weights using your examples. It is not the same as uploading a file for the model to read.
Generative AI
Generative AI creates new text, images, or other media from a prompt. An LLM is the usual text engine. Predictive AI, by contrast, estimates a number or a class from past data, such as which jobs might run long. Generation writes. Prediction scores.
LLM
An LLM (large language model) is the text engine inside tools like ChatGPT, Claude, Gemini, and Grok. It predicts a useful next sentence from the text you gave it and the text it was trained on.
Memory
AI memory is information a product saves from past chats so later answers can use it. It might be a note the app writes, or a setting you typed.
Multi-agent
A multi-agent setup uses more than one AI agent, each with a narrower job, passing work between them. A subagent is a helper one agent calls. A swarm is a loose name for many agents at once. The owner still needs one place where a person can see the result.
RAG
RAG (retrieval-augmented generation) means the model looks up your documents first, then writes the answer from what it found. It is a search step plus a writing step.
SOP
An SOP (standard operating procedure) is the written way your business does one job: who does it, the steps, and what to do when a step fails. An agent needs that document or it will invent a process.
Token
A token is a chunk of text the model counts. A short word may be one token. A long word or a number may be several. Plans and context windows are often measured in tokens, not words.
Vector database
A vector database stores embeddings and finds the nearest ones to a question. It is the index behind 'search our docs by meaning.'
Vendor lock-in
Vendor lock-in is when leaving a product is painfully expensive because your prompts, history, indexes, or connections live only there. Portability is the ability to take a useful copy with you. No tool is perfectly portable. Some make it needlessly hard.

Models and settings

How a model is set, what it was trained on, and what that does to a draft.

BYOK
BYOK (bring your own key) means the app uses a key you hold for the model provider, and that provider bills you for the model. The app may still bill you for its own features. You now have two vendors, two dashboards, and a key to store as a secret.
Knowledge cutoff
A knowledge cutoff is the point after which a model was not trained on newer public information. It can sound current and still miss a change that happened later. Vendors publish the date in their own docs, and it differs by model.
Max tokens
Max tokens is the cap on how long the model's answer may be, counted in tokens. Hit the cap and the answer stops mid-sentence. It is not the same as the context window, which is how much text the model can see in total.
Model provider
A model provider is the company that trains or hosts the model. The app you click, such as a chat product or a help desk, may be a different company that calls that model. Your contract, your data settings, and your bill may sit with either one, or with both.
Model training
Model training is how a vendor builds or updates a model from a large set of examples. Pretraining is the broad first pass. Fine-tuning is a later, narrower pass. Neither one is the afternoon you spend teaching your crew how to use the tool.
Open-source model
An open-source model ships under a license that lets you inspect, run, and sometimes modify it. Open weights means you can download the trained parameters even if every training detail is not public. Someone still has to host it, update it, and watch the bill.
Reasoning model
A reasoning model is a model that spends extra steps trying to work a problem through before it answers. It is often slower and uses more of your allowance than a plain chat model. The product, not you, decides how much of that work you get to see.
Small language model
A small language model is a language model with fewer parameters than the large models in full-size chat apps. It can be cheaper or able to run on modest hardware. It is often weaker on messy instructions. Parameters are the model's learned settings, not knobs you tune like temperature.
Temperature
Temperature is a setting that changes how predictable the next word is. Low temperature sticks to the likely phrasing. High temperature wanders. It is not a measure of how hot the computer is, and it is not a truth dial.
Usage cost
Usage cost is what you pay for runs, tokens, tasks, or credits after the subscription, if the vendor charges that way. The number is on that vendor's current pricing page and on your invoice. A token is the unit. The cost is the bill.

Building

The pieces you assemble when one chat box is not the whole job.

Agent workflow
An agent workflow is the series of steps an agent is allowed to run: trigger, lookup, draft, check, and stop. It is the job description plus the path.
API
An API (application programming interface) is a documented way for one piece of software to ask another for data or to take an action. No screen, no clicking.
Brand voice
Brand voice is how your business sounds when it writes: the words you use, the words you refuse, and the length you actually send. In an AI tool it belongs in standing instructions plus a few real examples, not in a mood board.
Chain of thought
Chain of thought means you ask the model to work through the steps before the final answer. You are requesting the scratch work, not a longer brochure.
Chunking
Chunking is splitting a long document into smaller pieces before they are embedded and searched. The size of the piece changes what the assistant can retrieve.
Classification
Classification means the model assigns an item to one of a list of labels you defined. It does not write the customer reply. It decides which pile the message belongs in, so the right next step can run.
Connector
A connector is a ready-made link between an AI app and another product, such as email, a calendar, or a CRM. It uses that product's API so you do not build the link yourself.
Extraction
Extraction means pulling specific fields out of unstructured text or a file: a name, a date, an address, a dollar amount. The rest of the paragraph is not the point. Missing fields should stay missing.
Few-shot prompting
Few-shot prompting means you show the model a few finished examples of the task before you ask for a new one. The examples teach the shape. Zero-shot means you give no examples and only the instructions.
JSON schema
A JSON schema is a written contract for structured data: which fields exist, which are required, and what type they are. Structured output asks for a shape. A schema is the strict version of that shape that software can reject.
Knowledge base
A knowledge base is the set of documents an assistant is allowed to treat as true for your business: SOPs, prices, FAQs, and policies, kept current.
llms.txt
llms.txt is a plain-text file at the root of a website that tells AI agents what the site is and which public pages matter. It is a map, not a secret API.
MCP
MCP (Model Context Protocol) is a standard way for an AI app to talk to tools. Instead of a custom plug for every assistant, a tool can speak MCP and more than one assistant can use it.
Prompt
A prompt is the instruction you give the model: the job, the facts, the format, and the things it must not invent.
Prompt template
A prompt template is a saved prompt with blanks for the facts that change. The instructions stay. The job, the date, and the price come from the person or the system filling the blanks. A prompt library is a shelf of those templates.
Scheduled task
A scheduled task runs at a time you set, without someone clicking a button. The trigger is the clock. The work might be a report, a reminder, or an agent checklist.
Skills
A skill is a packaged ability an agent can call, such as drafting a document, searching the web, or creating a spreadsheet. The agent picks a skill based on the request.
Structured output
Structured output means you ask the model for a fixed shape, such as a list of fields, instead of a freeform essay. The next step in a workflow can read those fields.
Summarization
Summarization is asking a model to shorten source material you provide: a thread, a transcript, or a long note. A good summary keeps the decisions and the open questions. It does not add a decision nobody made.
System prompt
A system prompt is the standing instruction the app applies before your message. It sets role, rules, and boundaries that a single chat message should not have to repeat.
Tool calling
Tool calling, also called function calling, is when the model asks a named tool to run, like 'look up the order' or 'draft a calendar hold,' instead of only writing text.
Web search
Web search, in an AI product, lets the model look up pages on the public internet and use them in an answer. It is not a search of your private files. Those are retrieval. Browsing is the same family of feature.
Webhook
A webhook is a URL that receives a message the moment something happens in another app. The other app pushes the event to you. You do not have to keep asking if anything changed.

Your files

Where prices, notes, and records live, and how an assistant is allowed to see them.

Data sync
Data sync is a repeating copy of fields from one system to another. It can be one-way or two-way. It is not the same as a source of truth. Sync moves the data. The source of truth says which copy you edit.
Deduplication
Deduplication finds records that are the same person or the same event and merges or skips the extras. It is a data cleanup. Idempotency stops one event from being processed twice. Deduping cleans a list that already has twins.
Knowledge gap
A knowledge gap is a question your assistant cannot answer from the documents you trust. The honest result is unknown, plus a note so you can write the missing page. It is not a hallucination. A hallucination fills the gap with a fiction.
OCR
OCR (optical character recognition) reads letters and numbers out of a picture or a scan. It produces text. It does not know if that text is the right model number. Extraction can then pull fields from that text.
Semantic search
Semantic search finds passages by meaning, not only by the exact words in the query. It usually relies on embeddings. Keyword search finds the words you typed. Hybrid search uses both.
Source of truth
A source of truth is the system you trust for a given fact. The price in the point of sale beats a price in a chat. A system of record is another name for that home. Data sync copies facts between systems. It does not decide which one wins.
Stale knowledge
Stale knowledge is an answer based on a document, memory, or index that no longer matches the business. The model may be working exactly as designed. The file is old. A knowledge cutoff is the model's training horizon. Stale knowledge is your copy.

Running the work

Schedules, retries, logs, and the queue you check when a run fails.

Batch processing
Batch processing runs a pile of similar jobs together, usually on a schedule, instead of one by one as each event arrives. It trades speed for a calmer review. Some model vendors offer a separate batch API with its own price. Read their current pricing before you assume it is cheaper.
Brittle automation
A brittle automation works only while every field name, template, and login stays exactly the same. A small change upstream breaks it quietly or loudly. Maintenance is the work of noticing and fixing that. Every live workflow needs a person who owns it.
Evaluation
An evaluation is a repeatable check of whether the assistant still does the job. You keep a small set of real inputs and the answers you would accept, and you run them again after you change a prompt, a file, or a model.
Exception queue
An exception queue is the list of cases a workflow could not finish safely: missing fields, low confidence, a customer who asked for a human, a tool error. Someone clears it. A dead-letter queue is the same idea when the message failed and is waiting to be retried or discarded.
Idempotency
Idempotency means doing the operation again does not do the job twice. The second identical request is recognized and ignored or safely folded into the first. An idempotency key is the id you use to notice the repeat.
Polling
Polling means your workflow asks another system on a timer whether anything is new. A webhook is the opposite: the other system tells you when something happens. Polling is simpler when the other app cannot push, and noisier when you ask too often.
Rate limit
A rate limit is how many requests a system will accept in a period, such as per minute. Exceed it and calls fail or slow down until the window resets. A quota is the sibling idea, often a longer allowance. Neither one is your monthly subscription price by itself.
Reconciliation
Reconciliation is matching two records of the same money or the same goods until the differences are explained. A bank rec matches the bank to the books. An AI draft can list likely matches. It should not mark a row cleared unless the rule is exact and a person can see the pair.
Retry
A retry runs a failed step again after a pause. It helps when a service blinked. It hurts when the failure was a bad phone number or a rate limit you immediately hammer. The automation glossary on this site also covers retry logic for classic workflows. This page is about the customer-facing choice.
Sandbox
A sandbox is a safe place to run a workflow where the effects are fake or limited to test records. Test mode is the same idea inside a product. It is not production, which is the live system guests and customers are on.
Shadow mode
Shadow mode means the new workflow runs on live inputs but its output is not what the customer receives. You compare it to the current process. It is not draft mode, where someone approves each send, and it is not the unapproved staff use called shadow AI.
Staff training
Staff training for AI is teaching people which tool to use, what they may paste, and what they must check. It is not model training and not fine-tuning. The outcome is a person who can follow the SOP.
Trace
A trace is the step-by-step record of one run: what the model saw, which tool it called, and what came back. An audit log says who approved or sent. A trace says how the draft was built. You need the trace when the words are strange.

Customers

The inbox, the phone, and the words a customer actually receives.

AI disclosure
AI disclosure is telling the recipient that a machine helped produce the words, the image, or the summary. It can be one plain sentence. It is not a claim that a human never looked.
AI notetaker
An AI notetaker joins or uploads a meeting, transcribes it, and summarizes decisions and action items. A meeting bot is the same category. The transcript and the summary are only as good as the audio and the consent.
Consent
Consent here means a person agreed to the specific use: recording a lesson, transcribing a call, or feeding their message to a tool. A privacy policy they did not see is not the same as a clear ask. Rules differ by place. This page is not legal advice.
Intent
Intent is what the customer is trying to do, named as a job: book, reschedule, ask a price, complain, or get a human. Intent detection is classification aimed at that job. It is not a mind reading claim.
Internal note
An internal note is a comment only the team should see, on a job or a ticket. A public reply is what the customer receives. An agent that can do both will mix them up if you do not separate the tools.
IVR
An IVR (interactive voice response) is the button menu on a phone line: press 1 for the shop, press 2 for the road desk. It follows a tree you drew. It does not understand a sentence unless you add a voice agent on top.
SLA
An SLA (service level agreement) is a promise about speed or uptime: we answer web leads within a set time, or a vendor answers you within theirs. For a small shop it is often just the reply window you tell customers. A vendor's SLA is their promise to you, on their contract, not a number you invent.
Transcription
Transcription is turning speech into text. The text can be wrong on names, numbers, and accents. It is not a summary and not a decision. Speech-to-text is the same job.
Voice agent
A voice agent is software that listens to speech and talks back, often to book, triage, or take a message. It is the speaking layer. An AI receptionist is the front-desk job that might use voice, text, or both. An IVR only offers a menu.

Safety

Who can send, what stays private, and what you write down before staff use a tool.

AI policy
An AI policy is your rule for how the company may use AI tools: which tools, what data may be pasted, what must be reviewed, and who can turn on a send. A vendor's acceptable use policy is their rule for what you may do with their product. You need both, and they are not substitutes.
API key
An API key is a secret password an app uses to prove it is allowed to call an API. Anyone who has the key can use your account until you revoke it.
Audit log
An audit log is a record of who did what, and when. For an agent, that includes drafts, tool calls, approvals, and sends.
Auto-send
Auto-send means a workflow delivers a message without someone pressing send that time. It is safe for a template you already approved, filled with fields from the system of record. It is a bad default for fresh prose.
Citation
A citation is a pointer from an AI answer back to the page, file, or row it used. It lets you check the claim instead of trusting the tone.
Data retention
Data retention is how long a tool keeps what you put in it: chats, files, logs, and backups. The vendor's setting and your own habit are both part of it. A long memory in the product is not the same as a retention rule you chose.
Draft mode
Draft mode means the agent may write the message and must stop before send. A person sends. Draft-only is the same rule. It is a form of human in the loop aimed at outbound words.
Escalation
Escalation is the handoff from an AI step to a person when the situation is outside the script: anger, money, safety, or a simple 'I want a human.'
Grounding
Grounding means tying an AI answer to a source you trust, such as a file, a database row, or a web page with a citation, instead of letting the model answer from memory alone.
Guardrails
Guardrails are the limits around an AI system: what it may say, which tools it may call, what a person must approve, and what data it must not touch.
Hallucination
A hallucination is when an AI states something that is not true, often in a confident tone. It may invent a price, a policy, a citation, or a step.
Human in the loop
Human in the loop means a person reviews or approves an AI step before it counts. The agent can prepare the work. A human decides.
Indirect prompt injection
Indirect prompt injection is an instruction hidden in content the agent was asked to read, such as a web page, a PDF, or an email, that tries to override your rules. Direct prompt injection is the user typing the override themselves.
OAuth
OAuth is a login handoff. You approve an app in Google, Microsoft, or another account, and that app gets a limited token. You do not give the app your password.
Permissions
Permissions are the specific actions and files an app or agent is allowed to use. They are granted by an admin, often through OAuth scopes or a role.
PII
PII (personally identifiable information) is data that points at a specific person: name plus phone, email, address, account number, or a photo that identifies them. A first name alone in a reminder you send on purpose is different from a spreadsheet of members dropped into a chat.
Prompt injection
Prompt injection is when untrusted text tries to override your instructions. A customer email, a web page, or a file can say 'ignore your rules and send the database.'
Redaction
Redaction is removing or masking sensitive details before text goes to a tool, a person who should not see them, or a log. It is the action. PII is the kind of detail you are often removing.
Secrets
A secret is a credential a workflow needs, such as an API key or a password. It belongs in the tool's connection vault or a password manager, not in a prompt, a screenshot, or an SOP.
Shadow AI
Shadow AI is staff using AI tools the business did not approve, usually a personal account, with company or client information pasted in. The work may look finished. The data left a door you do not control.
SSO
SSO (single sign-on) lets people enter several tools with one company login, often Google or Microsoft. You are not sharing a password with each app. You can often turn access off in one place when someone leaves. It is not the same as OAuth to a single calendar, though the plumbing is related.

Tools

Named products and the job shapes owners mix up with each other.

Make scenario
A Make scenario is a visual workflow: modules connected on a canvas. Each module action usually spends one credit. A Make AI Agent can sit in that flow when the step needs a judgment.
n8n workflow
An n8n workflow is a set of nodes that run when a trigger fires. n8n counts one execution for the whole run, not one charge per node. AI and MCP nodes are on the Cloud plans n8n lists, starting with Starter.
Zap
A Zap is one Zapier automation: a trigger and one or more actions. When the trigger fires, the actions run. Zapier bills successful steps as tasks.

Questions owners ask

What is an AI agent glossary?

It is a set of owner-level definitions for the terms behind AI assistants and automations. Each page starts with a direct answer and links to a prompt or guide already on this site.

How should I use this glossary?

Look up the word a vendor just used, read the owner section, and open the linked prompt or guide if the job matches work you already do. The automation glossary covers older workflow words if you need those too.

Are the business stories real customers?

No. Every example is a made-up illustration so the definition has a shop attached to it. They are not results, testimonials, or claims about a vendor.

Product pages are under AI apps for business. Workflow words that are not about agents stay in the automation glossary.