AI Foundations for the Workplace

AI in Plain English

The AI tools you'll meet at work are called "large language models", or LLMs for short. You don't need to use that term out loud. What they actually are: programs that have read an enormous amount of text (most of the public internet, millions of books, vast amounts of technical writing) and have gotten very good at predicting what words should come next.

That sounds small. It isn't. To be good at predicting words, the model has to "understand" grammar, context, facts, tone, and the shape of an answer. The result feels like talking to a knowledgeable colleague, except it doesn't get tired, doesn't get offended, and doesn't remember you between conversations (unless you turn that on).

You don't need to know how it works to use it well. You do need to know where it's reliable and where it isn't.

That second part is what this session is really about. The tool is less important than your mental model of when to trust it.

The Four Tools You'll Meet at Work

There are many AI tools. These are the four you'll actually bump into in an NT workplace.

Honourable mention: NotebookLM. A Google product, different job. You upload documents, it answers questions grounded in those documents only. Very good for making sense of a stack of papers, reports, or policy docs. Available as a free web tool.

Copilot vs ChatGPT, Claude, Gemini, What's Actually Different

They all look like chatbots. The difference that matters is what they can see and what happens to what you type.

Microsoft 365 Copilot (the paid one, built into Word, Excel, Outlook, Teams) can read the documents, emails, meetings, and files your account has access to. It answers grounded in your real work. Your prompts and the content it reads stay inside your organisation's Microsoft tenant, covered by Microsoft's enterprise agreements.

Consumer ChatGPT, Claude, Gemini can't see anything unless you paste or upload it. Anything you put in could (depending on settings and licence) be used to improve the product, be seen by a human reviewer, or be stored indefinitely. Those are third-party services, not part of your workplace tenant.

Enterprise or business versions of ChatGPT, Claude, Gemini exist with stronger data protections that start to resemble the Microsoft enterprise arrangement. If your organisation is licensed for one, you'll know.

Copilot sees your work. The consumer tools only see what you hand them. Both are useful. The difference decides what content each is safe for.

Where Your Data Goes

The single most important thing to understand about AI at work: anything you type or upload is data that leaves your keyboard and goes somewhere. The question isn't whether data moves, it's where, and what happens when it gets there.

We'll come back to this in detail in the Responsible Use, Privacy & Policy session. The mental model starts here.

What AI Is Good at in a Workplace

The tasks where AI earns its keep in an NT office, community organisation, or government team.

First drafts. Emails, briefing notes, meeting recaps, short reports, grant sections. You'll rewrite parts. The first draft saves the blank-page time.

Summarising. A long policy document, a meeting transcript, a long email thread, a 40-page PDF. "What are the key points and what's being asked of me?"

Rewriting for different audiences. Turning technical content into plain English. Adjusting tone (formal, direct, friendly). Making a long paragraph shorter.

Translation. Both between languages and between registers (policy jargon to plain English).

Research synthesis. "I've got five documents on this topic. What's the common thread? Where do they disagree?"

Structured thinking. Pros and cons lists. Risk tables. Decision frameworks. Questions you should ask before signing something.

Drafting formulas and data transformations. "Turn this messy list into a clean table." "Write an Excel formula that does X."

Tutoring. "Explain this to me like I haven't done this kind of work before." "Quiz me on this policy until I know it."

Meeting preparation. "Given these three documents, what questions should I expect?" "What background do I need to read before this conversation?"

Where AI Is Unreliable

The honest list of what AI gets wrong. Knowing this is the difference between a professional user and someone who gets caught out.

It hallucinates. Makes up facts, quotes, references, court cases, policy clauses, phone numbers. Writes them with total confidence. If the specifics matter, verify them against a primary source.

It's out of date. Most AIs are trained months or a year ago. They won't know last week's machinery-of-government change, yesterday's policy update, or a recent appointment, unless a live search feature is on.

It's weak on local specifics. It will be confidently wrong about a specific NT program, a particular community organisation, a local bylaw, or a named person. The more local and niche, the less to trust.

It struggles with arithmetic and counting. Bad at maths without a calculator tool. Will miscount words, dates, or items in a list. For anything numerical that matters, check.

It doesn't know what it doesn't know. Unlike a thoughtful human, it will confidently answer questions outside its training rather than saying "I don't know."

It can't see context you haven't given it. If you ask for a "short email" without saying who it's to, the tone will be generic. The AI fills in the blanks from averages, and averages usually aren't you.

The Prompting Skills That Work session covers how to prompt it in ways that reduce these failures. The Responsible Use session covers how to review output before it goes out.

The Human in the Loop

The single most important idea for using AI at work: the AI drafts, the human decides. That's not a legal disclaimer, it's how the tool is actually designed to work. Copilot included.

AI output is the starting point of your work, never the ending point.

What "human in the loop" looks like in practice:

Read before you send. Every AI-drafted email, memo, or paragraph gets a human read before it leaves you. Not a skim. A real read.

Fact-check the specifics. Dates, names, numbers, quotes, statutory references. If the AI said it, and it matters, check it.

Own the decision. The AI can lay out options. You choose. When something goes wrong, "Copilot wrote it" isn't a defence, your signature is on the work.

Keep your judgement sharp. Using AI for first drafts doesn't mean your writing skills atrophy. Read the AI's work as a critic, not a consumer. Disagree with it. Rewrite parts. That's the work.

Sensible Defaults for the Rest of the Course

The posture we'll use for the rest of the Working with AI strand. Not a policy, just a working stance.

Your working stance

• For work content, prefer Copilot (inside M365) over consumer tools.
• For your own research, writing, and learning, consumer tools are fine, just don't paste sensitive work content into them.
• Never put personal information about others, identifying details, or content you don't own into a free consumer AI.
• Treat AI output as a draft. Read before sending. Check specifics.
• When you're unsure, ask the AI: "how confident are you, and where should I verify?"
• Use AI most aggressively where the downside of getting it wrong is small (personal drafts, learning, brainstorming). Use it more carefully where the downside is large (external communications, policy, advice).