AI terms explained, and what each one means in a proposal

The basics

Every term below is defined in one or two sentences, and where the word carries a commercial consequence the definition says what that consequence is. Nothing here needs a technical background to follow, and if a definition uses another piece of jargon to explain itself, treat that as a fault in the definition rather than a gap in you.

Read it in whatever order is useful. If you are checking a proposal, the sections on automation, messaging and measurement are the ones that decide what you actually receive.

Model

A file of learned patterns that takes text in and produces text out. When a supplier says they are building AI for you, the model is usually the one part they did not build.

Large language model

A model trained on very large amounts of text to predict what comes next. That prediction is the whole mechanism, which is why a model can be fluent and wrong inside the same sentence.

Prompt

The instruction and context you send the model. In a working installation most of the effort goes into the prompt and the material attached to it, not into choosing a model.

System prompt

The standing instruction sitting above every conversation that sets the rules, the tone and the boundaries. This is where you write what the system must never say, promise or quote.

Token

The unit a model reads and writes, roughly a short word or a piece of one. Usage is counted in tokens, so a design that pastes an entire document into every request pays for that document every single time.

Context window

How much the model can hold in view at once, measured in tokens. A system that has to consider your whole product catalogue needs retrieval rather than a bigger prompt, and this is the reason.

Hallucination

A confident, fluent answer that is untrue. Better models produce fewer, but prediction can always do it, so grounding rules and a permitted way to say I do not know matter more than which model you picked.

Temperature

A setting controlling how varied the output is. Low is correct for anything customer facing that states policy or price, and a demonstration run at high temperature is showing you creative writing rather than your use case.

Inference

One run of the model to produce an answer. Training happens rarely; inference happens every time a customer sends a message, and it is what your ongoing usage is made of.

Multimodal

A model that can read more than text, usually images and audio too. That matters directly in this market, because customers send voice notes and screenshots of payments instead of typing.

Benchmark

A standard test used to compare models. Benchmark position rarely predicts how a system behaves on your own messy inputs, so treat it as a first filter and never as evidence.

Agents, tools and the words for what a system does

This group has drifted furthest from any settled meaning. Two suppliers can both say agent and mean things that share almost nothing, so when one of these words appears in a proposal, ask for the sentence underneath it.

Agent

A system given a goal that decides which steps to take and uses tools to take them. The word is used very loosely, and a lot of what is sold as an agent is a fixed workflow with a chat box in front of it.

Tool calling

The mechanism that lets a model do something outside the conversation, such as look up an order or create a record. Without it, a model can only talk.

Tool

One capability you expose to the model, described well enough that it knows when to reach for it. Those descriptions decide more of the behaviour than most buyers expect.

Orchestration

Deciding what runs, in what order, with what handed to what, and what happens when a step fails. This is where an installation is usually won or lost, because the model itself is the easy part.

Guardrail

A rule constraining what the system may say or do, enforced outside the model rather than requested inside the prompt. Asking a model politely not to quote a price is not a guardrail.

Human in the loop

A design where a named person approves or corrects particular actions before they take effect. Sensible on refunds, complaints, prices and anything a customer can later hold you to.

Escalation

The handover from the system to a person, with the conversation summarised rather than pasted. A system without a clean escalation path produces angrier customers than having no system at all.

Autonomy

How much the system does without being asked. Vendor material often uses fully autonomous to mean unsupervised, which describes a risk position rather than a capability.

Multi-agent

Several specialised agents passing work between them. Occasionally the right shape for a problem, and often a complicated answer to something one well described workflow had already solved.

Evals

A repeatable set of real inputs with known correct outcomes, run every time anything changes. Without evals you are editing a live system on the strength of how the last demonstration felt.

Knowledge, or how a system knows anything about your business

A model arrives knowing nothing about your prices, your stock, your policies or your customers. Every term in this group exists to solve that one problem, and the choice between them is usually the largest technical decision in a project.

Retrieval augmented generation

Usually shortened to RAG. Rather than relying on what the model absorbed during training, you look up the relevant passages from your own material and put them in front of it before it answers.

Embedding

A numeric representation of a piece of text that sits near other text with similar meaning. This is what lets a search find the right paragraph when the customer used none of your words.

Vector database

The store that holds embeddings and finds the closest matches quickly. For a small knowledge base an ordinary database with good search is often enough, so treat a vector store as a decision to justify rather than a default.

Chunking

Splitting documents into passages small enough to retrieve usefully. Bad chunking is the commonest reason a retrieval system returns half an answer, and it stays invisible until somebody reads the retrieved passages.

Grounding

Requiring the answer to come from retrieved source material instead of the model's memory. The practical test is whether the system can show you which document a sentence came from.

Citation

The reference back to the source of an answer. Worth insisting on internally even if a customer never sees it, because it is how you work out why a wrong answer happened.

Fine-tuning

Further training of a model on your own examples so that it adopts a format or a style. It teaches behaviour rather than facts, and using it to install knowledge is a common and costly mistake.

Training data

The material a model learned from. Your operational documents are not in it, which is why any question about your own business needs retrieval of some kind before the model answers.

Knowledge base

The organised set of documents the system is allowed to answer from. Whoever keeps it current owns the accuracy of the whole installation, and that person should be named before go-live rather than after the first wrong answer.

Hybrid search

Combining literal keyword matching with meaning based matching. Useful here because product codes, brand names and part numbers have to match exactly while the question wrapped around them does not.

Automation and the plumbing underneath it

Most of what a business buys under the name AI is really this: workflows, connections between systems, and careful handling of the cases that do not fit. These words are older than the current wave and they are the ones that determine whether an installation keeps running.

Workflow

A defined sequence of steps that runs when something happens. Most business automation is this rather than anything agentic, and for most processes that is the correct choice.

Trigger

The event that starts a workflow. A message arriving, a form being submitted, a file landing in a folder, or a time of day.

Node

One step inside a workflow in tools like n8n, Make or Zapier. A node does a single thing: read, transform, decide or write.

Webhook

A URL that one system calls the instant something happens, so the other system never has to keep asking. Faster and cheaper than polling, and it needs a plan for what happens when the call fails.

API

The published way one piece of software talks to another. If a system you depend on has no API, automating around it means screen scraping or re-entry by hand, and you want to know which before committing.

Integration

A working connection between two systems, including the awkward parts: authentication, field mapping, error handling, and what to do when a record exists in one place and not the other.

Idempotency

The property that running a step twice has the same effect as running it once. Without it, a retry can send a second invoice to a customer, and retries do happen.

Retry and backoff

Attempting a failed step again after an increasing wait. A network hiccup at three in the morning should not need a person, and a permanent failure should not be retried forever.

Exception handling

What the system does when something does not fit the rules. Demonstrations skip this part, and it is the part that decides whether an installation survives its third week.

Queue

A holding line for work waiting to be processed. It stops a sudden burst of messages from either overwhelming a downstream system or being quietly dropped.

Self-hosted

Running the software on infrastructure you control rather than the vendor's. n8n can be self-hosted, while Make and Zapier cannot, which matters when the data moving through a workflow is not allowed to leave your own environment.

Audit log

A record of what the system did, when it did it, and on whose behalf. If you cannot answer six weeks later why a particular customer received a particular message, you do not have one.

Messaging and measurement

Two groups that get skipped in proposals and then cause the arguments. The messaging terms are platform rules you do not get to negotiate. The measurement terms are the ones a supplier will use to tell you the work succeeded, so it is worth agreeing what each one means while you are still friendly.

Messaging, and the WhatsApp words that trip people up

WhatsApp Business API

The programmatic route into WhatsApp used by businesses handling real volume. It is a different thing from the WhatsApp Business app, and the rules attached to it shape what any automation can and cannot do.

Template message

A message the business starts, using wording approved in advance. You cannot write a new one at the moment you need it, so every template you are going to need has to be planned before launch.

Session message

A free-form reply sent inside the window a customer's message opened. This is where ordinary back and forth conversation happens.

Messaging window

The period after the customer's last reply during which the business can send free-form messages. It closes, and once closed the business needs an approved template before it can speak again.

Opt-in

The customer's recorded permission to be messaged. Required, and the record of how and when it was given is exactly what you need on the day a complaint arrives.

Business Solution Provider

The intermediary, usually shortened to BSP, through which most businesses reach the WhatsApp Business API. Ask who yours would be and whose name the account sits in before signing anything.

Handover

Moving a live conversation from the automation to a named person with the context attached. On a number shared by several staff it also means recording who picked it up, otherwise nobody can say who promised what.

Broadcast

Sending an approved template to many opted-in contacts at once. Governed by the same opt-in and template rules as everything else, and the quickest way to get a number reported if the list is poor.

Measurement, including the numbers used loosely

Baseline

What the number was before you changed anything. Measure it for a few weeks before installation, because after go-live there is no honest way to reconstruct it.

First response time

How long a customer waits for the first meaningful reply. The easiest number to move with automation, and the one owners here feel most directly, because an enquiry that sits for hours is often an order that went somewhere else.

Containment rate

The share of conversations the system finishes without a person. Only meaningful next to a quality check, because a system that ends conversations by exhausting people scores well on it.

Deflection

Usually a vendor word for the same idea as containment, framed from the support team's side of the desk. Ask precisely how it is calculated before accepting it as a result.

Resolution rate

The share of conversations where the customer's actual problem was solved. Harder to measure than containment and worth considerably more.

Escalation rate

How often the system hands a conversation to a person. A rising escalation rate is information rather than failure, and it usually points at a gap in the knowledge base.

Cycle time

How long a whole process takes from end to end, from the enquiry arriving to the thing being done. Automating one step while the queue in front of it stays the same moves nothing a customer can feel.

Payback

The point at which time or cost recovered covers what the installation took to build. Ask a supplier which measurement they would accept as proof, and write it down before work starts.

What a glossary cannot settle for you

Knowing the vocabulary does not tell you whether a piece of work has the right shape for your business, and it is a poor substitute for the two questions that matter: who operates this after it is live, and what do you own at the end. A buyer who has memorised this page can still sign a proposal that leaves the accounts in the supplier's name.

These definitions also describe common usage rather than a standard. There is no authority that fixes what agent or deflection mean, vendors stretch both, and a supplier who defines a term differently is not necessarily wrong. Ask them to write down their version and put it in the contract, which settles it in a way an argument about definitions never will.

If you came here to price a project, this is the wrong page. Vocabulary does not scope work. The scope comes from watching one process for a week and counting how often it repeats, which no amount of reading does for you.

Common questions

What is the difference between an AI agent and an automation?

An automation follows a sequence you defined in advance: when this happens, do these steps in this order. An agent is given a goal and chooses its own steps using the tools you gave it. Most business processes want the automation, because a process with rules that fit on one page does not benefit from a system that improvises.

Do I need to understand any of this to buy AI work?

You need enough to ask three things: where the system's knowledge comes from, what happens when it does not understand, and whose name the accounts sit in. Everything else you can reasonably delegate. A supplier who cannot answer those three in plain language is telling you something.

What is RAG in plain English?

RAG means looking things up before answering. Instead of hoping the model remembers your policy, the system searches your own documents, pulls out the relevant passages and puts them in front of the model with the question. That is why a RAG system can cite the document it answered from, and why keeping those documents current becomes somebody's job.

Is fine-tuning how I teach a model about my business?

No, and this is one of the more expensive misunderstandings in the field. Fine-tuning teaches a model how to behave, such as always replying in a particular format. Facts about your prices, stock and policies belong in retrieval, because they change and retraining every time they change is not practical.

Which of these terms are mostly marketing?

Fully autonomous usually means unsupervised, which is a description of who carries the risk. Deflection is a support metric with a flattering name. Multi-agent is often a complicated way of describing a workflow. None of them are dishonest words, but each one hides a question worth asking directly.

What does the context window actually limit?

It limits how much text the model can consider in a single request, including your instructions, the conversation so far and anything you attached. When your material is larger than the window, the answer is retrieval rather than a larger window, because paying to send an entire catalogue with every message is both slow and worse at finding the right line.