You belong in the future.

This was the most common thread of the whole night, so if it was you, you were in very good company. Start with one task, not a transformation. Pick the thing you do every week that you'd happily never do again, and a lot of you named yours at the stations: emails, quotes, reports, admin. Take just that one thing and, for a week, do it with AI alongside you, asking it to draft, simplify or speed up each step.

You don't need a grand plan, and "going in blind" is exactly the right way to walk in. The Context Kit on the hub (your Business Profile and your Ideal Client Profile) is the highest-leverage starting point there is. Once AI knows who you are and who you serve, everything else gets easier. Then Larissa's free Demystifying AI module picks up from there at your own pace. You don't need a strategy yet. You need one small win.

For most small businesses you don't need a special tool. Claude or ChatGPT will do a lot here. Export your data from the system as a CSV or spreadsheet, upload it if the tool lets you, and ask in plain English: "summarise jobs by month," "which clients are most profitable," "draft a one-page report from this." Claude is particularly strong at working through a spreadsheet and showing its reasoning.

Two notes. Strip or anonymise anything personal before uploading (see the privacy answers below), and always sanity-check the numbers on anything that matters, because AI can make confident arithmetic mistakes. If you outgrow that, the next step is connecting the system directly, something like Power BI or a proper integration, which is more of a strategy-session conversation.

Good catch. Claude (or ChatGPT) doesn't actually log in and post by itself. What it does brilliantly is create the content: captions, hooks, a month of posts from a single idea, in your voice. To get it published, you connect it to something that does the posting. Either you copy and paste into your scheduler (Metricool, Buffer, Later, Meta's own planner), or you wire it up with an automation tool like Zapier or Make so a generated post flows into your scheduler automatically.

Newer versions can connect to some tools more directly, but for now the simple version is: AI writes it, your scheduler posts it. Honestly, copy and paste into Metricool is the simplest reliable setup, and where we'd start.

Yes, the tech exists. Tools like HeyGen and Synthesia build a video "avatar" from a sample of you, and others can clone your voice, so you type a script and "you" delivers it. It genuinely works and people use it for repetitive content.

The honest caveat: audiences can increasingly tell, and the thing that makes your content land is usually the real, slightly awkward, human you. Our take is to use an avatar for the boring, repeatable stuff (FAQs, course intros, updates) where polish matters more than presence, and keep your actual face for the moments that build trust. And before you outsource your face entirely, let AI do the heavy lifting around the video first: scripts, teleprompter, editing. That alone makes the awkward bits a lot less awkward.

We held this one on the night because it's a rabbit hole of its own, but here's the short version: yes, and it's real. You can now describe a simple tool in plain English, a booking form, a little calculator, a way to sort some data, and AI will build a working version with no code from you. It's early, it isn't magic, and the more involved the thing gets the more it helps to know what you're doing. But the floor has dropped right out of what it costs to turn an idea into something that actually works.

The honest advice: get solid on the everyday stuff first, the context kit, using it as a daily assistant, then come and find one of us when you want to go down this path. It deserves a proper conversation, not a paragraph.

Both, and knowing which is the real skill. If the answer is close but off in tone or detail, keep going in the same chat: "make it warmer," "shorter," "now do it for this client." The AI remembers the context and builds on it. If it's gone genuinely off track, wrong direction, confused, talking in circles, start a fresh chat with a better opening prompt. Once a conversation is muddled, it tends to stay muddled.

Rule of thumb: refine to improve, restart to redirect. And the best fix is almost always a stronger first prompt: who it's for, what you want, what "good" looks like, and a bit of context, which is exactly what the Context Kit gives you on tap.

On tools, there are four worth knowing, and they mostly do the same thing. ChatGPT is the popular all-rounder. Claude is known for lovely, natural writing, and it's what we used for the live demo. Gemini is Google's, and it's great if you live in Gmail and Docs. Copilot is Microsoft's, built into Word, Excel and Outlook. Our advice: pick one, get good at it, and the skill carries across.

On getting better answers out, that's the whole "context" thing from the night. The quality of what comes back is mostly down to the quality of what you put in: who it's for, what you want, and enough background about you and your business that it isn't left guessing. The Context Kit on the hub is built to do exactly that.

Two parts. On tokens: AI reads and writes in "tokens," little chunks of text, roughly three-quarters of a word each, so a short sentence is often around 15 to 20 tokens. It matters for two reasons. The tools are charged by tokens behind the scenes, and each model can only hold so many tokens at once, its "memory" for a conversation, which is why very long chats sometimes start forgetting the beginning.

On free: the big tools (Claude, ChatGPT, Gemini) all have genuinely useful free tiers, and we expect those to stick around, because they're how you get hooked. But free comes with limits: fewer messages, the older or smaller models, and a wall on busy days. If AI becomes part of how you work, the paid tier, around NZ$30-ish a month at current prices, pays for itself fast in time saved. Think of free as "try everything," and paid as "now it's a tool I rely on."

The instinct is to teach "how to use the tools," but those change every year. The skills that last are the human ones AI doesn't have: asking good questions, judging whether an answer is actually true, thinking critically, creativity, and knowing what you actually want in the first place. Curiosity and the confidence to learn new things beats any single tool.

Practically: let them use AI as a thinking partner and a tutor, but keep them doing the hard parts themselves, writing, maths, reading real books, so they build the judgement to spot when AI is wrong. The kids who thrive won't be the ones who can prompt. They'll be the ones who can think, and who aren't afraid of the technology.

Honestly, anyone who tells you precisely is guessing, including the people building it. The realistic direction: AI keeps getting better at tasks, writing, summarising, analysis, coding, customer support, design drafts, so more of those get automated. But a job is a bundle of tasks, judgement and relationships, and those are much harder to replace wholesale.

The pattern so far is that AI changes jobs more than it deletes them. The person who uses AI well tends to replace the person who refuses to. Most exposed: routine, screen-based, repetitive work. Most protected: anything built on people, trust, hands and judgement. Which is exactly why we'd rather you were in the room learning it than waiting to see.

Anyone who gives you a precise picture is guessing, so take this as direction, not a forecast. But the direction is fairly clear. The biggest shift is that AI stops being a separate thing you go and open. It gets baked into the tools you already use, Outlook, Teams, Excel, your internal systems, so it's just there in the background, the way spellcheck is now. The "go to ChatGPT" step quietly disappears.

Day to day, your job tilts from doing the task to directing and checking it. Less typing from scratch, more "draft this, pull that together, summarise these," then you review and decide. For somewhere data-heavy like Northpower, the near-term wins are the obvious ones: meeting notes and actions captured automatically, reports drafted straight from the data, and being able to ask your own systems a question in plain English instead of hunting through them. The one thing we'd bet on: the people who do well over that 2 to 5 years won't be the ones who waited to see how it shook out. They'll be the ones already in the habit now.

We'll be straight with you: there isn't yet a clean, trustworthy number, and anyone quoting a precise figure should be treated with caution. It's too new, and layoffs get blamed on AI for lots of reasons. The most credible source to follow is the World Economic Forum's Future of Jobs Report, which surveys employers worldwide on what they expect to automate versus create.

In short: meaningful churn is coming, some roles shrink, new ones appear, but the apocalyptic headlines are running well ahead of the actual data. We'll keep pointing you to real reporting rather than scary numbers.

Here's the thing: there isn't one clean qualification that turns into an AI job, and anyone selling you a single "AI paper" as the answer is overpromising. The field moves faster than the courses can keep up, so formal quals are always a step behind.

It also splits two ways, and they're different paths. One is becoming AI-capable in the work you already do, which is what most "AI roles" in business actually mean right now, and what we'd point most people toward. The other is becoming a technical specialist, the data and machine-learning end, which is a real career change and a much longer road. Worth knowing which one you're reaching for before you spend any money.

For the first, the fastest route isn't a degree. It's getting genuinely good with the tools and being able to show you've applied them to real problems in your field. That's what gets you hired or promoted, and there are plenty of solid free starting points from Google, Microsoft and the AI companies themselves. If you do want something structured and NZ-based, AcademyEX (Frances Valintine's) is the one we know firsthand. Lian did her Master of Technological Futures there, and they run shorter micro-credentials too, not just the full degree. Good place to look first.

Yes, and it's where most of the real value is, with one boundary. The whole reason the Context Kit matters is that AI gets far more useful when it knows your business: your services, your tone, your processes, your strategy. That's the good stuff, and for most small businesses the risk in sharing it is low.

The boundary is other people's information and anything legally protected (see the next two answers). One setting to check: consumer tools sometimes use your chats to improve the model unless you turn it off. Claude and ChatGPT both let you opt out, and their business tiers don't train on your data at all. So feed it everything about you freely, and be deliberate about anything that isn't yours to share.

Simple rule: don't paste anything you'd be uncomfortable seeing leaked or used to train a model. Concretely, keep out passwords, banking and card details, client personal or identifying information without consent, health or other sensitive records, anything under an NDA, and whole databases of customer data. Your own business strategy, processes and content? Go for it.

If you genuinely need AI to work with sensitive data, that's a job for a business or enterprise tier with a no-training agreement, not your personal free account. Turn off chat history and training in the privacy settings as a baseline. And remember, under NZ's Privacy Act 2020 you're responsible for the personal information you hold, so putting it into a random AI tool can be a breach, because you're still responsible for where that information goes. When in doubt, leave it out.

Two halves, and both are covered in detail in this section and the one above, so briefly. On security: the rule is don't put anything in that you'd be uncomfortable seeing leaked, passwords, client details, anything sensitive or under an NDA, and the answers above lay out the specifics.

On authenticity: this is the flip side of context. Generic AI "slop" happens when the AI has nothing of you to work from. The fix is feeding it your voice, your real material, the things that make you you, so what comes back sounds like you wrote it. Authenticity isn't something AI takes from you. It's something you protect by staying in the driver's seat and giving it enough of yourself to work with.

This is exactly the situation where you go slow and start away from the sensitive data, and yes, there's still real room. Two principles.

First, separate the risky from the safe. Never put personally identifiable or sensitive client information into general AI tools, full stop. But a huge amount of your work isn't that: internal drafting, plain-English-ing dense policy, first drafts of reports, templates, summarising your own meeting notes. Start AI there, where no individual's data is involved. You'll get most of the time-saving with none of the exposure.

Second, where AI might ever touch real data, that's an organisational decision, not a personal-account one. You'd want tools with the right agreements (data not used for training, hosted appropriately, signed off against your contracts and the Privacy Act). On the client side, the test is: "would this person be comfortable knowing exactly how their information was handled?" If you can't answer a clear yes, keep it human. Honestly, for an organisation like yours the right first step is a proper sit-down to map what's safe versus not, which is precisely what a strategy session is for.

The honest answer starts with what to keep human. Connection is your actual product. The relationship, the trust, the read on a probationer in a hard moment, none of that gets handed to AI, and you wouldn't want it to. So the rule for a business like yours is that AI works on the scaffolding around the mahi, never the mahi itself.

Where it earns its place is the business you run on top of the supervision. Turning your own frameworks and competency standards into plain-English resources. Drafting session-prep templates and reflective prompts. Writing the marketing that explains what you do. Summarising your own notes into something usable. That's hours back, with none of the people in it.

The hard boundary, and it matters more in your field than almost any other: never put identifiable or confidential supervision content into a general AI tool. Probationers, clinical detail, anything sensitive, that stays out. If you ever want AI working closer to real records, that's an organisational decision with the right tools and agreements, not a personal account, and it's worth a proper sit-down. Shannon's question just above covers very similar ground in more detail.

It does have a real footprint, and we won't wave that away. The data centres that run these models use a lot of electricity, and water to keep the servers cool. A single text question is genuinely small on its own. The part that adds up is the sheer scale, billions of uses a day, and the data centres themselves, which draw serious power and water, often in places already short on both. Heavier uses like generating images and video cost a lot more than a line of text, so what you reach for matters.

Two things help. Be intentional, use AI where it genuinely saves you real work rather than reflexively for everything. And read the people doing the real reporting. Karen Hao's Empire of AI (linked in the Deep Dive section below) digs into the human and resource costs, and Dr Karaitiana Taiuru's kaupapa Māori framework, he tangata, he karetao, he ātārangi, a person, a puppet, a shadow, is a grounded way to hold both the usefulness and the cost at the same time. Several of you raised this at the stations too, so you're not alone in sitting with it.

Great question, and there's a real thing underneath it. Two worries usually get mixed together. One is "model collapse," the fear that as the internet fills with AI-generated content, models trained on that slop start to degrade. It's a genuine research concern, but the labs are very aware of it and are leaning on curated and licensed data, not just scraping the open web.

The other is quality sources, news, research, universities, healthcare, increasingly blocking or charging for their data. That part is true, and it's why companies are now signing licensing deals to keep access. So far, each generation of model is getting more capable, not less. But your instinct is sharp: the quality of what goes in absolutely shapes what comes out, which is the whole reason we banged on about "context" all night.

Yes, very much, and you're right that it's worth naming while it's early. These models are trained mostly on English-language, US-and-Europe-heavy internet, so the "default" voice, examples, spelling and assumptions skew that way, and te ao Māori, Pasifika and even plain NZ context are underrepresented. You'll see it in American spelling, northern-hemisphere seasons, US examples, and a flattening of anything outside the mainstream.

The fix on your end is context: tell it you're in New Zealand, who your audience is, the reo or kupu you use, your spelling, and it adjusts well. The deeper issue, whose knowledge is in the model at all, is exactly what Dr Taiuru's framework (linked in the Deep Dive section below) speaks to. Treat the default output as a first draft with a worldview already baked in, not a neutral truth.

Not the end, but a real shift. People increasingly get their answer straight from AI, Google's AI Overviews, ChatGPT, Perplexity, without clicking through to a website. So "rank number one on Google" matters a little less, and "be the source the AI quotes" matters more. People have even started calling this AEO or GEO, answer or generative engine optimisation.

The good news is the fundamentals overlap. Clear, genuinely helpful, well-structured content about what you actually do still wins, because that's what the AI pulls from too. AdWords isn't dead, but expect click-throughs to soften over time. This is exactly the kind of thing Lian digs into in a website audit.

This is exactly what the live demo was built to show. Same tool, same task, two very different results depending on the context you give it. For listings and open homes, the move is to build your context once, who you are, your patch, your voice, the kind of buyer you want, and then every post, every listing blurb, every follow-up comes out sounding like you and pitched at the right people, in a fraction of the time. The Jess example in the slides, a real estate agent and a Kamo listing, walks through precisely this, the lazy version versus the good one. Start there.

This is a brilliant, very modern problem. AI flattens nuance. It gives the "average" answer and struggles with the holistic, the unpublished, the genuinely different thing you do. Two moves.

First, make your own material the best source on you. AI mostly reflects what's publicly written, so if your website, profiles and content clearly explain your philosophy, your nuance and your evidence, you shape what the AI says about your field, and about you. A thin or absent online presence just leaves the AI to fill the gap with generic or competitor framing.

Second, when a client arrives with a flat AI answer, treat it like any other objection: "That's the textbook version, here's what it misses, and here's why." It's an opening to show exactly the depth AI can't. The businesses that lose here are the invisible ones. The ones that win have a clear, well-documented point of view online.

Katie from Rumours Cafe asked how a small cafe could create a month of social content in a few hours while keeping it local and authentic. That question became our live demo on the night, so rather than answer it twice, the slides on the hub walk through exactly how it was done.