Sorry, what's an agent?

Agent is one of the newer words in AI, and it tends to arrive mid-conversation: in a software update, a vendor demonstration or a question from a partner. Token, context window and a file ending in .md are often close behind.

People in firms pick these words up as they go, from colleagues and from the tools themselves. This glossary covers them in 20 entries, in plain English, with what each one means for the work of a firm.

1How familiar the words are

In the Ipsos AI Monitor 2026, 69% of Australians said they have a good understanding of what artificial intelligence is. Fewer, 45%, said they know which types of products and services use it. Ipsos surveyed about 1,000 Australians aged 16 to 74 in March and April 2026.

For AI Adoption in Australia, a study the ANU ran with Google, researchers asked 3,577 adults in October 2025 how well they understand generative AI. Of those surveyed, 6.6% said very well and 24.6% moderately well. Another 45.7% had a basic understanding, 19.6% had heard of it without understanding it, and 3.5% had never heard of it.

In research by Deloitte Access Economics for RMIT Online, 7% of more than 2,000 Australian workers reached an advanced level of AI literacy and more than half were beginners, even though 84% had used at least one AI tool.

In the 2025 Victorian Lawyer Census, the lawyers not yet using AI showed more interest in training on how AI works (65%) than those already using it (53%). The report covers 1,739 respondents.

The National AI Centre's adoption tracker for December 2025 to February 2026 found agentic AI "largely untapped" among small and medium businesses, "reflecting both a lack of awareness and a lack of confidence in understanding how these tools work". The tracker surveys at least 400 business owners and decision-makers in each wave.

2Why shared words matter

What we tend to see is that a new word gets a nod in the meeting and a search afterwards, if there is time. A decision about a new feature can then rest with the one or two people who follow the terms, and the rest of the team hears about it later.

With a shared vocabulary, the practice manager and the IT provider can talk about which connectors are switched on in the same terms. A fee earner can say a draft was grounded in the matter file, and the partner reviewing it knows which source material the tool was asked to use.

New terms arrive with each software update, faster than one person can keep up with alone. A glossary the firm keeps, and the team adds to, holds that learning in one place. Building a Confident AI Team looks at where the wider confidence comes from.

3The tools

Generative AI. AI that creates new content. The Digital Transformation Agency describes generative AI as tools that generate content, such as text, images, software code, audio or video, "based on patterns learned from large volumes of data". A first draft of a letter of advice, a summary of a long contract and a tidied file note are all generative AI at work.

Large language model (LLM). The kind of generative AI that works with words. The OAIC describes it as "a type of generative AI that specialises in the generation of human-like text". It learns patterns from very large amounts of text and predicts what is likely to come next, a word or part of a word at a time. Large language models underpin the language capabilities of tools such as ChatGPT, Claude, Gemini and Microsoft Copilot.

Model. The engine underneath an AI tool. Providers release new models regularly, and one tool can offer several, often named with a version number. Two models can give different answers to the same question, which is one reason a firm may name the model as well as the tool in its guidance.

Chatbot and AI assistant. A chatbot is an AI tool you interact with conversationally, by typing or speaking. An AI assistant is an AI tool designed to help with tasks through conversation. Some are standalone, while others are built into software people already use, such as Copilot in Word and Outlook. Our post on Copilot licences looks at getting more from one.

Machine learning. An approach to AI in which systems learn patterns from data. The Tax Practitioners Board describes it as computers that "learn how to tackle problems and discover solutions independently, often by using artificial neural networks". The spam filter on the firm's email is a long-standing example.

Multimodal. A tool that can take in more than one kind of input, such as text, images, audio or a scanned PDF. A multimodal tool can read a photographed receipt or work from the recording of a client meeting, within what the firm allows into it.

4Working with the tools

Prompt. What you type or say to the tool. In the Digital Transformation Agency's words, "Your input is called a prompt, and the AI's reply is the output." A prompt can be one line, or a page with the matter background, the firm's tone and an example of a good file note. A shared prompt library keeps the ones that work.

Token. AI models process text in small units called tokens. A token can be a whole word, part of a word, a character or punctuation. Tokens are how the length of a conversation is measured, and how some tools are priced, which comes up when reviewing AI subscriptions.

Context window. How much the tool can hold in view at once: the prompt, any attached documents and the conversation so far, counted in tokens. A long matter file may not fit in full, and in a long conversation the tool can lose track of what was said at the start.

Hallucination. When a tool produces something that sounds right and is wrong. The Law Society of NSW gives one form of it: a tool producing "a reference that appears legitimate, but is in fact fictitious", such as a case that does not exist. The OAIC links it to the way generative AI predicts its output "based on likelihood" and the way it "tokenises input". Checking AI output before it reaches a client is covered in Are you the human in the loop?

Grounding and RAG. Grounding means giving an AI model relevant source material to base its answer on, such as a matter file, the firm's precedents or current search results. Retrieval-augmented generation, or RAG, is one way of grounding an answer. The system first retrieves relevant material from a larger collection, then gives that material to the model when it writes the response.

Reasoning model. A model designed to use additional internal reasoning before it produces its answer. Some tools offer it as a "thinking" option. It can help with harder problems that have several parts, such as checking a set of figures against each other, and the reasoning itself may not be shown to the user.

Training data. The material a model learned from. Some AI tools can also use what people type to train future versions, depending on the product, the type of account and its settings. Before client information goes into a tool, check which applies.

5Agents and connections

AI agent. An AI tool that carries out a task over several steps, deciding what to do next as it goes. The National AI Centre's Guidance for AI Adoption defines an AI agent as an "automated entity that senses and responds to its environment and takes actions to achieve goals". Asked to prepare for a client meeting, an agent might search the matter file, pull the latest correspondence, draft an agenda and suggest a time, checking back with a person at the points the firm has set. ASD's Australian Cyber Security Centre describes agentic AI as systems that rely on an AI model "to interpret and reason about the state of the world, make decisions and take actions". Because an agent takes actions, decide what it can reach, and where a person approves a step, before it is switched on.

API. Short for application programming interface: the way one piece of software asks another for information or to do something. The Australian Government's API site describes an API as "a tool that allows developers and product teams to re-use parts of existing systems". When a practice management system sends invoices to the accounting software, an API is likely to be doing the work in the background.

Connectors and MCP. A connector is a link that lets an AI tool access another system, such as email, document management or practice management software. MCP, the Model Context Protocol, is one way of building those connections. It is an open standard for connecting AI applications to external data and tools. MCP started at Anthropic, and in December 2025 it became one of the founding projects of the Linux Foundation's Agentic AI Foundation. Once a connector is switched on, the AI tool may be able to access data or take actions in that system, subject to the permissions it is given, so switching one on is something the firm and its IT provider decide together.

6Files and data

Markdown (.md). A plain-text file with a few symbols for formatting: # for a heading, asterisks for bold, a dash for a bullet point. John Gruber created it in 2004 so people could "write using an easy-to-read, easy-to-write plain text format" (Daring Fireball). AI tools often write in Markdown, which is why a pasted answer can arrive scattered with asterisks and hash marks. A .md file opens in any text editor and can be turned into a Word document.

JSON. A way of storing information as labelled pairs that another program can read, such as "client": "Example Pty Ltd" and "matter": "2026-114". The IETF standard describes JSON as "a lightweight, text-based, language-independent data interchange format". It turns up when data is exported from a system, or when an AI tool is asked to return results in a set structure.

CSV. A spreadsheet saved as plain text, with a comma between each column. Most systems can export one and Excel opens it. The IETF's RFC 4180 documents the format, which it says has been used "for exchanging and converting data between various spreadsheet programs for quite some time". A time-entry report exported as a CSV is something an AI tool can summarise or check.

SQL. Short for Structured Query Language, the language used to ask a database for information, such as every matter one partner opened this financial year. Practice management and accounting systems hold their data in databases, and many reports are built with SQL underneath. The language is set out in an international standard, ISO/IEC 9075. Some AI tools can write SQL from a question asked in plain English.

7When the words are shared

A firm glossary lets people check what a term means and how the firm uses it. It can sit beside any approved tool list and form part of induction.

8Starting a firm glossary

9Where we come in

Our investigation asks every person in the firm how they use AI, which tools they are comfortable with and where they want more support. It measures how long the firm's work takes today, so you start from a baseline of your own. The findings belong to your firm whether or not we go any further together.

AI Foundations comes next for the whole firm, followed by cohort workshops grouped by team or skill level.

If these words are coming up in your firm, reach out to our team. The first conversation works out whether your firm is the right fit for the AI Uplift Program, and whether we are the right people to help.

Building a shared AI vocabulary