Getting your head around AI: a plain English guide

We regard ourselves as relatively AI literate here at Creative HQ.

Working in innovation means it’s important to us to support the start ups, government agencies and corporates we work with to use the most recent technology, and to understand its benefits, opportunities and challenges. Not only that, but we love to experiment with our own ways of working too. How we can use the tech to do better work and free up time for the parts of the job that actually need a person, while steering clear of the usual change traps along the way.

Despite all that, it can feel like someone’s having a conversation in a new language when they talk about AI. Tokens, chips, LLMs. The people who are really embedded in the AI world sometimes don’t realise the rest of us are still getting our heads around it all.

So we thought it might be helpful to put together a plain English guide to the terminology, the technology, and the key ideas that sit underneath it.

How does it actually work?

If you’ve ever read that these tools “just predict the next word” and thought that can’t possibly explain what they do, you’re not alone. It’s true, and it still feels like it shouldn’t work.

A large language model is trained by being shown enormous amounts of text and asked, over and over, to guess what comes next. At the start it guesses badly. Every time it’s wrong, its internal settings get nudged a little. Do that billions of times across most of the written internet, and to get good at guessing the next word the model has to build an internal feel for grammar, facts, tone, and the shape of an argument. Guessing is the task. The understanding of patterns is what it has to develop to do the task well.

So when you ask it something, it isn’t looking up an answer in a database. It’s generating the most likely continuation, shaped by everything it picked up in training. That’s why it can write in your style, summarise a document it’s never seen, or sketch out a plan. Those are all ‘pattern completion’ that, at a scale, starts to look a lot like thinking.

It’s also why it can be confidently wrong. The same machinery that produces a fluent, sensible answer will just as happily produce a fluent, sensible answer that’s completely made up, because it’s working out what would plausibly come next, not what’s actually true. It’s a brilliant first drafter, or improver and a poor final authority.

The AI word directory (AKA “the words that get thrown around”)

The fundamentals of AI

Start with the basic one. LLM just means a large language model, the kind of system sitting behind ChatGPT, Copilot, Claude and Gemini.

A few words turn up the moment you start using one. A prompt is simply the instruction you give it. The context window is how much it can pay attention to at once, meaning your prompt, any documents you’ve added, and its own reply, all counted together in chunks called tokens (a token is roughly a word, or part of one). It’s why very long documents sometimes have to be broken up. They don’t all fit in the window at once.

The word that matters most for anyone thinking about using AI is hallucination. That’s when the model produces something fluent and completely confident that also happens to be wrong: a made up statistic, a court case that never happened, a policy that doesn’t exist. It isn’t lying. It’s doing exactly what we described earlier, generating a plausible continuation, and plausible isn’t the same as true. The fix you’ll hear about is grounding, which means pointing the model at your own trusted documents so its answers come from real source material rather than thin air.

Then there’s the word of the moment, which you’ll hear over and over this year: agent, or agentic AI. The shift it describes is from a tool that answers a question to one that can carry out a whole task. An ordinary chatbot tells you how to do something. An agent can plan the steps, use other software, and actually do it, with a person overseeing it. That’s what sits behind Copilot’s newer “agents”.

Two more travel alongside it. RAG, or retrieval augmented generation, is really just the technical name for grounding, fetching the right documents before it answers. And MCP, the Model Context Protocol, is a recently agreed common standard for plugging AI tools into other software and data sources. It works a bit like a standard adapter: rather than every connection between an AI assistant and your systems being built from scratch, MCP lets them talk in an agreed way. It’s dry plumbing, but it’s a big part of why agents can suddenly do useful things across different systems, and it’s why you’ll keep hearing the acronym.

If you ever want to look under the bonnet, a model’s size gets described in parameters, the billions of internal numbers it settled on during training, and the whole thing runs on specialised chips. Worth knowing they exist, but not something you’ll need in a Tuesday meeting.

Open or closed, self hosting, enterprise and sovereignty

It’s worth getting straight what “open” actually means here, because it catches people out. The “weights” of a model are, in plain terms, the trained model itself, the enormous set of numbers it ended up with once training finished. A closed model is one where the company keeps those weights to itself and only lets you use the model through its own service, like ChatGPT, Copilot or Claude. You never hold the model, you just get access to it. An open weight model is one the maker has released for anyone to download and run on their own computers, with Meta’s Llama models the best known example. The thing to watch is that “open” doesn’t automatically mean “free to do whatever you like with.” The licence still sets out whether you can use it commercially, which is why open weights and licences always come up together.

An enterprise agreement changes the deal. Your data stays inside your organisation’s own tenant and isn’t used to train anyone’s model, and you get the admin controls, security and compliance on top. That, rather than any extra feature, is a big reason organisations pay for it instead of letting staff quietly use the free version.

AI sovereignty is that same worry scaled all the way up to a country. It’s the question of how much real control New Zealand has over the AI it comes to depend on. Where is the data processed and stored? Can things run onshore? Are we locked into a small handful of overseas providers? For New Zealand, there is also the important consideration of Māori data sovereignty, and who gets to set the terms in the first place. Across much of the New Zealand public sector the working AI strategy has become “use Copilot, because it comes with the Microsoft licence we already hold.” That procurement led default is raising real questions about vendor lock in, even as a handful of agencies quietly test the alternatives. The sovereignty conversation isn’t happening in a policy paper somewhere, instead it’s happening in everyday purchasing decisions. AI sovereignty has recently felt a lot more topical, with the sudden, forced suspension of Anthropic’s Claude Fable 5 and Mythos 5 sparked by a government directive restricting access by foreign nationals due to a narrow software vulnerability/jailbreak.

Using it well at work, beyond redrafting emails

The trap most of us fall into is treating these tools as a faster typewriter. Drafting an email, tidying a paragraph, summarising a meeting. Useful, but it barely scratches what’s possible, and it’s where most people get stuck.

A quick word of honesty first, because the hype tends to skip this. Not everything is a win. A 2025 UK government evaluation found that asking AI to generate images, manage scheduling, or build PowerPoint slides didn’t actually save people any time, while reviewing code, summarising research, drafting documents and transcribing meetings clearly did. So yes, look at slides and spreadsheets, but the gains sit in the thinking and the first draft, not in asking it to assemble the deck for you. Australia’s Tax Office, for what it’s worth, found some of its biggest time savings in data visualisation, which is the spreadsheet angle done properly.

A rich example sits right on our doorstep: Hutt City Council

The team at Hutt City Council has been using AI across the organisation since late 2024, and they’ve gone well past drafting. They’ve rolled out around 300 secure licences across ChatGPT and Copilot, and built a set of custom tools they call AI Assistants for specific, repetitive jobs. Their list includes assistants that turn meeting transcripts into minutes, peer review memos heading to decision makers, vet building consents and Warrant of Fitness renewals, review traffic management plans, and pull common themes out of consultation data. They’ve put hard numbers on some of it too, like invoicing automation that saves a few minutes on each of up to a hundred invoices a week, and AI that turned a pile of handwritten City Summit submissions into a full report in two days, a job that would otherwise have taken weeks.

The part worth copying isn’t the tools themselves though. It’s how they set it up. Hutt City didn’t keep AI as something a central digital team does to everyone else. They train their own staff to build their own Assistants for their own needs, and they’ve even built an Assistant whose only job is to help people make new ones. That’s a shift, from “AI is a specialist function over there” to “the specialist capability lives inside each team.” It’s the same instinct as putting a technical person inside a delivery team rather than off in a separate unit.

It’s also worth knowing where Copilot itself is heading, because it changes the picture. The newer Copilot “agents” can connect to your organisation’s own data and take whole processes off your hands, rather than just answering a question in the moment.

None of this works without the boring scaffolding, though. Hutt City pairs all of it with an AI risk framework, a governance group, a public register of where they use AI, and a firm rule that a person always checks what the AI produces.

Where to now

The value isn’t in better prompts for your emails. It’s in finding the specific, repetitive, well bounded jobs in your own team, the ones that eat hours and don’t really need judgement, and pointing the tools at those with a person kept firmly in the loop. Start small, keep the guardrails on and tell your colleagues what’s working.