Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Last week we spent two days with Tourism New Zealand as part of a larger engagement to move them from single-player AI to multiplayer AI by training events teams, BDMs and reporting teams to build Copilot Agents and redesign shared workflows.
The biggest insight was how much confidence people gained once they understood AI’s ceiling.
For example, most non-technical workers have never been taught when AI is the right tool: it can interpret messy information, but it’s not the best fit for running data calculations or reliably retrieving system data, which often needs code, rules or integrations.
Once that distinction clicked, ambitious ideas became practical builds, and people could see what belonged in Copilot, what needed to sit elsewhere, and how the pieces of the AI puzzle fit rather than resorting to AI for everything.
If you’re looking to enable your people and redesign how happens across the business through the use of AI, get in touch.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
The Human Rights Commission says Te Tiriti must sit at the centre of how AI is governed. Its report names transparency, accountability, Māori participation and digital infrastructure as the four pillars, and Paul Goldsmith's office confirmed a response will be tabled in Parliament.
4 min read
Our take: The report asks for participation rather than consultation, not the same thing. Consultation means iwi get shown a finished design; participation means they are in the room while it is being decided. Being a Crown entity, this can’t be dismissed.
The Tenancy Tribunal issued AI guidelines after a tenant filed 316 apparently AI-written pages. The claim sought $40,000 over drinking water and was awarded $80 for a dryer, and six in ten property managers now report tenants using AI in tribunal matters.
3 min read
Our take: The Employment Relations Authority, insurance disputes and construction adjudication all run on the same model of written submissions read by a human. Every one of them should be writing its guidance now rather than after its own 316-page filing arrives.
MYOB found 64% of NZ businesses with 20 or more staff use AI, against 30% of those with one to five. Overall SME adoption reached 36%, up from 32%, and 92% of medium businesses said they would feel it immediately if AI use stopped.
3 min read
Our take: Only 30% of firms with one to five staff use AI, and the reason is boring. Bigger teams have someone who stumbles across a tool and mentions it at lunch, and a team of five is too busy to stumble across anything.
Christopher Luxon asked a Rotorua business audience how many were using Claude or Claude Code. The Spinoff fed his full remarks back to Claude, which described his style as confident, ideologically consistent and empathetically thin.
4 min read
Our take: Telling small business owners to adopt AI as their number one focus has a different impact when the same government has no AI policy of its own. The advice is not wrong, it is just being given by a mouthpiece rather than someone close to the technology.
A tech expert told RNZ that AI's effects are visible in New Zealand's youth unemployment numbers. Unemployment hit an 11-year high of 5.6% in the June quarter, with graduates named as the group struggling most to find entry-level work.
3 min read
Our take: Australia's government looked at the same question in July and found jobs for 20-24 year olds growing faster than for everyone older. A Stanford payroll study found the opposite in the US. Without significant data in NZ, it’s difficult to draw any conclusion on AI’s impact on unemployment yet.
Rotokawa School won national recognition for a bot that writes lesson plans for Years 1 to 6. It draws on 16 stories gifted by Ngāti Uenukukopako under a memorandum of understanding signed in 2018, the only Bay of Plenty school recognised this year.
3 min read
Our take: A small school with no AI budget produced something most orgs or the government would spend months or years procuring. Constraint tends to force clarity about what the tool actually has to do.
📚️ Mike’s Takes From The Week
Helping leaders and teams learn, adapt, and scale with AI.
1️⃣ OpenAI analysed 800,000 work messages, and the concerning number is not the one being reported: The headline was 43.5% of role-specific use crossing into another role's job. Underneath it, 61.5% of all messages were generic, tied to no role at all. Emails, summaries, tidying notes. If no role owns the task, nor does the person in the middle.
Read
2️⃣ Twelve hours with the Tourism New Zealand team, and the blocker was never the technology: Every person started with an ambitious idea, and 95% of those builds were achievable in Claude with minimal uplift. Finding the ceiling of a Copilot licence made people more confident, not less. Building one agent surfaced terms and supplier rules the business had never written down.
Read
3️⃣ Individual AI wins are becoming table stakes, because every competitor can buy the same licences: Six levels sit above the ceiling, from task augmentation to organisational redesign. The P&L only shifts at level three, where half the old steps stop existing because they only existed to move work between people. Watch revenue per head, because flat means tools got bought and nothing else changed.
Read
4️⃣ Two decisions separate the rollouts that work, and neither of them is the training programme: Move people out of prompting alone into teams redesigning work together, then redesign the work itself one workflow at a time. Neither happens without AI engineering talent in the room, which is why 90% of firms experiment and under 10% show P&L returns.
Read
🎙️ The AI Corner Podcast
This week's guest is Jackson Lee, Chapter Lead for AI Engineering at Air New Zealand. Hear:
How Air New Zealand embeds AI engineers directly into business teams to help redesign work and build AI capability alongside domain experts.
Why the role of the software engineer is shifting from writing code towards architecture, review, judgement and working effectively with AI coding agents.
How Air New Zealand uses strong engineering fundamentals, eval-driven development and cross-functional governance to move AI safely into production.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Skill and plugin security scanning entered beta, letting Enterprise customers run automated checks on third-party skills and plugins for malicious content. (Aug 6)
Self-hosted environments for Claude Code entered public beta, running sessions on a customer's own infrastructure with internal network access and compliance controls. (Aug 6)
Managed Agents gained per-session spend caps that pause work, a mid-turn advisor model, region control for inference, and automatic skill loading from a repository. (Aug 7)
Claude Code 2.1.224 added self-hosted runners, plugin installs from HTTPS zip with SHA-256 pinning, cross-machine messaging between sessions, and removed the 200-subagent cap. (Aug 7)
Claude Code 2.1.225 added gateway spend-limit warnings naming the cap and reset time, plus a workspace trust prompt for agents. (Aug 8)
Anthropic cut false positives in Fable 5's biology safeguards by roughly 85%, opening more legitimate biology and healthcare work while keeping dual-use protections. (Aug 7)
OpenAI
GPT-5.6 Luna became the default model for Free and Go users, message limits on text chat were removed, and a Think button was added. (Aug 6)
Luna API pricing fell roughly 80%, alongside the expanded free-tier access. (Aug 6)
GPT-5.6 Sol gained a reliability update with a thought slider for controlling reasoning depth. (Aug 6)
ChatGPT Voice gained file uploads mid-conversation and Projects context, referencing recent project chats, sources and instructions. (Aug 7)
The Atlas browser stopped working, with its agentic browsing folded into the ChatGPT desktop app and Codex; bookmarks, tabs and history do not migrate. (Aug 9)
Pasted content over 10,000 characters now converts to an attachment automatically instead of filling the ChatGPT composer. (Aug 4)
OpenAI paused Astra activities failing strengthened security controls after tests could not rule out Critical-tier cyber capability, meaning autonomous zero-day discovery in hardened systems. (Aug 8)
OpenAI acquired presentation startup NextSlide, with the whole team joining to work on ChatGPT presentation and visual document features. (Aug 8)
Ask Maps went global and gained opt-in Personal Intelligence, starting with Gmail so it can factor in upcoming flights and reservations. (Aug 6)
Gemini trip planning now builds itineraries from live Google Flights and Hotels inventory combined with a user's own stated preferences. (Aug 6)
Gemini Enterprise took custom MCP server data stores to general availability for secure access to private data and custom tools. (Aug 7)
Workspace Studio automation can now push text, Drive links, YouTube and web URLs into a Gemini Notebook as sources without manual adding. (Aug 7)
Demis Hassabis moved from DeepMind CEO to chairman of DeepMind and chief scientist of Alphabet, stepping back from daily operations. (Aug 5)
Open-Weight Model Updates
The White House framework exempts open-weight models from federal security review entirely, applying the 30-day pre-release window only to closed frontier models. (Aug 4)
Mistral released Shieldstral, a 3B multimodal safety classifier under Apache 2.0 that takes plain-language policies at inference time and runs on a single 16GB GPU. (Aug 4)
An OpenRouter snapshot put open-weight models at 69.1% of token volume against 30.9% closed, with Chinese models taking eight of the top ten. (Aug 5)
Microsoft
Kimi K3 reached general availability in GitHub Copilot, off by default for Business and Enterprise until an admin enables it. (Aug 6)
GitHub Copilot weekly release added multiple concurrent sessions in a Sessions sidebar, with Auto now showing which model handled each request. (Aug 7)
The Copilot impact dashboard gained a return on investment section alongside usage metrics. (Aug 7)
Developers can now set the reasoning level for the Copilot cloud agent, and trigger automations from an issue or pull request comment. (Aug 3)
Microsoft 365 E3 discounts extended to 31 December 2026 while all E7 promotional offers retire on 1 October. (Aug 3)
Microsoft began imposing per-engineer AI token budgets internally, making model spend a managed line item. (Aug 5)
NVIDIA
The Open Secure AI Alliance behind it passed 120 members a week after forming with roughly two dozen. (Aug 4)
Meta
Meta launched Muse Code in beta, a terminal coding agent built on its Muse Spark model that spawns parallel sub-agents in isolated worktrees. (Aug 5)
Meta guided to $130 to $145 billion of 2026 capital spending against a 91% drop in free cash flow. (Aug 5)
xAI
Grok Imagine Image 2.0 shipped as Quality Mode on web and mobile with region-level editing, background removal and multi-reference editing across five inputs. (Aug 7)
Amazon
Bedrock launched Web Search for OpenAI GPT models, grounding answers in current web results without data leaving the customer's AWS environment. (Aug 4)
Amazon Cognito became a skill in the Agent Toolkit, letting coding agents configure and troubleshoot authentication through packaged workflows. (Aug 7)
Apple
The Gemini-powered Siri reached the iOS 27 consumer beta, delivering personal context and on-screen awareness with full release expected in September. (Aug 3)
Perplexity
Projects gained shared files and a shared Brain, Personal Computer shipped on Windows, and Model Council added side-by-side model comparison. (Aug 4)
🤦 ️ AI Fail Of The Week
We all love AI, but it’s certainly far from perfect 🤔 …
✨A few people have asked…
It’s Mike here, I run The AI Corner.
I'm not just into writing about AI. I run Allexive. We work with businesses that have spent on AI licences and training and still can't point at what it changed. We redesign the work so it does.
👋 Mike & Erin