Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Last week, I asked readers to send me examples of work they thought AI should help with, but were struggling to make work. Thanks to everyone who responded (welcome to send more through).
Two examples stood out:
Checking an installation against engineering standards
A plumber wants to describe an existing installation or proposed job and ask: “Does this comply with the current standard?”. The answer must use the correct version of the standard and point to the exact clause. A general AI chatbot cannot reliably do this when standards are restricted, updated regularly and unavailable in its approved source material.
Processing large legal documents
A legal professional wanted AI to examine a 300-page evidence file, identify where each individual document began and produce an index with the correct page numbers. They also wanted it to extract events from a 100-page witness statement, place them in date order and combine repeated information without losing the original paragraph references. The AI struggled because chatbots are designed to generate language. They are not reliable page counters, filing systems or document-processing programs.
My recommendation for both was similar:
Let AI read language and make limited judgements.
Let ordinary software handle page numbers, calculations, sorting and version control.
Connect every answer to the page, paragraph or standards clause it came from.
Have a person check and approve the result.
For the legal example, this could become a reusable Codex or Claude Code Skill. It would process every page, ask AI to identify document boundaries, build the index and create a numbered PDF.
The engineering example would need an authorised library containing the correct versions of the standards. AI could understand the plumber’s question, while fixed rules and calculations would perform the compliance checks.
This is a common reason AI projects fail: the whole job is handed to a chatbot, even when much of it is really an automation or software problem. AI can help design and create that software, then handle the parts that require understanding language.
I recently covered this approach with three students in my Advanced AI Operator Coaching Programme. They learnt how to use Claude Code and Codex to build reliable tools around their own work, rather than relying on a chat window to do everything.
If you would like to learn more about the programme, reply and let me know.
One of the better uses of AI video. Famous memes all wrapped into one.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
New Zealand’s new online safety bill bans under-16s from social media and AI companion services. Platforms must infer age from behaviour or facial scans, with fines up to 10% of global revenue, and Labour supplied the votes after ACT and NZ First refused support.
3 min read
Our take: Notionally the bill makes sense, but it is hard to see how it gets implemented, because making Big Tech pay for these protections has been a losing game everywhere it has been tried. A rule from Wellington also shifts responsibility away from families, teens and the companies themselves. A society that relies on the Government for every guardrail will create a behaviour where people stop taking accountability. I find that worrying longer term.
The FMA warns AI can fabricate pay slips and bank statements convincing enough for home loan fraud. The warning follows Australian authorities uncovering potentially hundreds of millions of dollars in suspected mortgage fraud, and the regulator wants lenders to lift their detection.
3 min read
Our take: This headline frustrates me. The headline is intentionally pessimistic and skips the other side of the coin. Sure, AI can be used for nefarious reasons. It also gives lenders better fraud detection than they have ever had. The same technology that fakes a pay slip can also spot a fake pay slip much faster than a human can. We also have new technologies and regs, such as Open Banking, to reduce this risk by replacing documents with verified bank-feed data, and AI on top of it can tell customers when they are actually eligible before they apply. These problems and opportunities exist in any industry where approvals run on PDFs and documents. The institutions that adapt their process will be safe in the long run, and the ones that do not will feel the brunt of nefarious actors.
Auckland student Caden Scott’s startup Azonic raised NZ$1m to build an AI analyst for police. Blackbird led the round the same week it closed Fund VI at A$1.05b, the largest venture fund raised in Australasia, with agencies in NZ, Australia and the US already using Azonic.
3 min read
Our take: Some will argue the frontier models could handle this, but an AI analyst reading police case files is a narrow, high-stakes product opportunity that big vendors will avoid. Also, the geographical moat these platforms build is huge, because once an agency runs on one, mountains of case context pile up inside it. That makes identifying a better alternative harder every month, and building in-house is unlikely without the expertise.
NZAero restarted CT-4 Airtrainer production after 14 years, and wants AI reading its flight data. CEO Stephen Burrows says glass cockpits and expanded data collection open the door to AI partnerships on pilot training and safety, targeting the global pilot shortage.
3 min read
Our take: The volume of flights today produces data that could massively level up training capability in a travel market that is not going anywhere but up. Delivering that information is not the hard part, because the systems exist and just need updating. The hard part is making sense of the data and converting it into a curriculum purpose-built for the trainee, the environment and the airline. The key is NZAero’s domain expertise applied alongside AI experts inside the build. Outsourcing that pairing is the biggest risk to the concept staying viable.
MPI says AI could add $15 billion of value across the primary sector’s supply chain. Chief insight officer Jarred Mair says the technology removes the trade-off between productivity and sustainability, and predicts humanoid robots on farms within a decade.
3 min read
Our take: Agritech opportunities are abundant for NZ, and Ireland and the Netherlands bring similar primary-industry strengths to the same race. When the technology is democratised for anyone, access to the tools stops being the advantage. The differentiators become entrepreneurship and the willingness to try and fail, government support and investment, and industry focus that fosters our early-stage advantages. Without those, NZ stays a taker of technology rather than a maker in a space where we have an advantage.
Researchers propose using AI to find the last “hold-out” possums in New Zealand’s wilderness. The Conversation piece describes acoustic sensors and detection models that locate surviving pests after control operations, the hardest and costliest phase of eradication.
4 min read
Our take: Conservation is becoming one of NZ’s most practical AI use cases, with listening networks and camera traps producing data no team of rangers could review. The tools are cheap and the labour they might replace is scarce.
Tech NZ says NZ ranks third in APEC for generative AI use, but firms lack the digital basics. Its report “Digital Foundations First” recommends connectivity, cloud systems, cyber security and usable data before broader AI adoption across SMEs.
3 min read
Our take: The report says the constraint on SME AI is not the AI, it is patchy connectivity, missing cloud systems and light security infrastructure. Fixing the boring layer first is slower to announce and faster to pay back.
A NZ$4.6m programme will scan apples’ internal quality with hyperspectral imaging, no cutting required. The Crown puts in NZ$1.84m through the Primary Sector Growth Fund, targeting 80% detection accuracy by mid-2027 and a working packhouse prototype by 2028.
3 min read
Our take: none, just news.
Northlanders range from enthusiastic to petrified about AI, and many workers feel “AI guilt” using it. The Northern Advocate reports experts see the bigger business challenge as using AI safely and effectively rather than whether to use it at all.
5 min read
Our take: AI guilt means workers hide productivity gains from managers, so the business never learns what is working. Leaders are the ones who can change this narrative and help everyone benefit from individual learnings.
Stop typing what you could say in 10 seconds.
Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.
📚️ Mike’s Takes From The Week
Helping leaders and teams learn, adapt, and scale with AI.
1️⃣ The people getting reliable results with AI are not relying on the model alone: Sceptics who say AI invents sources are testing one component, the model on its own. Business-grade output comes from the system built around it.
Read
2️⃣ If OpenAI’s CEO still copy-pastes between tools, AI adoption will take longer than the hype suggests: Sam Altman admits he has not fully made the shift. Getting one person productive with AI is one thing. Getting 100 people there with the right access, reliable information, and agreed ways of working is organisational change, and it needs a plan.
Watch
3️⃣ “What roles do we need to hire?” is the last question in AI workforce planning, not the first: A workforce redesign lead’s sequence: get clear on outcomes, map the actual work as verbs, decide what AI takes on, reverse-engineer the skills, then choose build, hire, borrow, or AI.
Read
4️⃣ Counting agents is a misleading measure of AI progress: A dedicated agent pays off when a workflow repeats often, is shared across a team, stays stable for months, and is measurable. For one person, the better build is the system around AI.
Read
5️⃣ AI agents now use 5x more tokens than humans, and that changes who a business serves: One question for an executive team: what happens when the main visitor to a website is an AI screening services on a customer’s behalf, against systems built for humans?
Read
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Memory now works across chat and Cowork, with remembered information viewable and editable under Topics plus a control for sensitive topics. (Aug 25)
Claude Cowork gained a built-in Chromium browser in its side panel, letting Claude navigate, click and fill forms without an extension. (Aug 26)
Claude Code 2.1.243 to 2.1.251 added a curated model picker, keyless Console sign-in, a restricted mode, model-switch hooks and a spend-limit bar. (Aug 28)
Sony Music Publishing, Warner Chappell and other publishers sued Anthropic and two co-founders, alleging illegal torrenting and scraping of copyrighted compositions. (Aug 28)
Anthropic is reportedly pitching IPO investors a $30 trillion addressable market ahead of raising up to $100 billion at a $2 trillion valuation. (Aug 27)
OpenAI
The ChatGPT browser extension expanded to Edge, Brave, Opera and Vivaldi, with Site tools letting agents work directly with websites via WebMCP. (Aug 25)
Scheduled tasks gained webhook triggers from Gmail, Slack and GitHub events, and Free users got three active tasks. (Aug 25)
Users can connect multiple Google accounts for Gmail, Calendar and Contacts in a single conversation. (Aug 28)
The DALL-E GPT retired from ChatGPT, with users told to download images and move to ChatGPT Images. (Aug 30)
OpenAI will cut Cursor's direct model access from 12 November, saying it cannot trust SpaceX to honour its terms of service. (Aug 29)
Gemini 3.5 Transcribe reached general availability with real-time streaming, speaker diarisation, word timestamps and language detection across 85-plus languages. (Aug 26)
Gemini Live became an agentic voice assistant with spoken daily briefings, hands-free inbox management and handoff of multi-step jobs to the Spark agent. (Aug 26)
Gemini-based data classification in Google Drive entered open beta, letting admins auto-label files from plain-language instructions for DLP and retention policies. (Aug 28)
Chinese Frontier Labs
Zhipu confirmed the stealth OX Alpha model was GLM-5.3-Flash, a natively multimodal 320B model, and released its weights under MIT. (Aug 27)
Z.ai then published full GLM-5.3 weights (753B, 1M context), swapping MIT for a licence requiring security review above $10 billion revenue. (Aug 28)
Alibaba open-sourced Qwen3.8-Flash-Next, a 125B multimodal preview of the Qwen4 architecture claiming Qwen3.7-Plus performance at one-ninth the training cost. (Aug 26)
DeepSeek's roughly $7 billion round at a $74 billion pre-money valuation neared close, with a Shanghai STAR Market listing in preparation. (Aug 26)
NVIDIA
NVIDIA reportedly agreed to buy Hugging Face for $12.9 billion, its largest acquisition, though no signed agreement has been confirmed. (Aug 27)
Microsoft
Anthropic's Sonnet 5 became available and default for higher-reasoning tasks in Word's Copilot models menu. (Aug 25)
Edit with Copilot in Excel now executes Python for advanced analysis, writing results back into the workbook. (Aug 25)
✨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