Notes and takeaways from the Aotearoa AI Summit 2026

The Aotearoa AI Summit ran over two days at Tākina in Wellington on 8 and 9 September, organised by the AI Forum New Zealand. Day one covered the infrastructure and capability the country needs, the trust and social licence it has to earn, and the gap between pilots and production, and had the four Aotearoa AI Hackathon finalists pitching live. Day two turned to the workforce: what AI does to skills, entry-level roles and how work is designed, and then to the economic opportunity and the case for building for export.

These are our notes, session by session, for anyone who couldn’t make it or wants the takeaways in one place. We’ve kept them faithful to what each speaker said and added our own commentary throughout; look for the “Our take” blocks. Two caveats: we arrived part way through a couple of sessions, and we missed the day-two panel on the productivity dividend (Hema Sridhar from Koi Tū and Dr Viveca Pavon-Harr, moderated by Carol Hirschfeld), so it isn’t covered here.

The short version

It’s a long post. If you only have five minutes, these are the threads that ran through both days, with links to the sessions behind them.

• Adoption is near universal; transformation isn’t. 91% of New Zealand organisations use AI and 4% have changed how they operate because of it. The gap is made of evals nobody wrote, workflows nobody redesigned and costs nobody tracked. Greg Davidson and the pilot to production panel.
• Cheaper per token, dearer per task. Agentic work uses around a thousand times the tokens of a chat. Measure cost per successful outcome. Greg Davidson and Accenture.
• Human in the loop is not enough on its own. A reviewer who can’t override, or never does, is ceremony. Greg Davidson.
• Sovereignty means governing your dependencies, not owning the stack. Exit rights, model portability and knowing where your data, keys and logs sit. Sovereign capability panel.
• Small, community-driven AI works. Te Hiku Media trained a world-class te reo Māori model on one GPU with data its community chose to give. Peter-Lucas Jones.
• Trust is earned, and only a third of New Zealanders currently extend it. Govern AI as infrastructure, embed Te Tiriti from the start, and treat participation as different from consultation. Human Rights Commission, earning trust panel and TVNZ.
• Redesign the work, don’t automate the old process. Electricity only paid off once the factory was rebuilt around it. Accenture and Kiwibank.
• The entry-level pipeline is a leadership decision. If AI takes the tasks graduates learned on, you have to build a new way to grow judgment. Augmenting Human Potential panel and Maria Mingallon.
• Small is a superpower, and New Zealand should build for export. Thirty-day pilots with money as the metric, led from the top. John-Daniel Trask and David Steele.

Reuben Davidson: what is the role of AI in Aotearoa going forward?

Labour’s spokesperson for economic development, science, technology and innovation, and broadcasting, media and the creative economy sat down with Will Mace from the NBR for the opening conversation, a day before the party released its AI action plan, so he was careful not to pre-empt it. A few things still came through clearly.

• Data centres are the question voters ask. Two or three times a week someone raises them at a street-corner meeting or in his office, usually about water and power. His response is to explain why the country has them and needs more, while insisting that operators take responsibility for their environmental impact. The “thousands of jobs” story does the sector a disservice; once a data centre is built, it doesn’t employ thousands.
• The public conversation is missing. Government may be doing work on AI, but “a little too quietly”, which gets in the way of a bipartisan approach and leaves the public out of understanding both the risks and the opportunities. He pointed to the Human Rights Commission’s point that process matters as much as outcome.
• Small business readiness. Labour has committed to a funding package covering a new initiative, an extension of the existing pilot programme, and support for resilience. The range of what “using AI” means in small business is wide: some are running agents; some have “bought a laptop with Copilot on it”. A question from the floor suggested Digital Boost, which reached 65,000 businesses, as the model. His answer: considered.
• Light touch is not the right approach, but nor is chasing every headline. His preferred move is to elevate where AI oversight sits within government. On public service efficiency, AI should help sort the data so humans decide better, not replace the human in the loop; he was unimpressed by ministers announcing headcount cuts “because AI” without costing the AI.
• Bias for Māori in AI applications? Asked what policy steps are being taken, his answer was blunt: none he knew of from the current government.
• Efficiency versus value. Asked how New Zealand shifts from a race for cost saving to taking the positive opportunities, he agreed the framing is wrong and said it comes down to the priority government places on AI across every portfolio, not just the tech one.

Martijn Verburg: the age of intelligence

Martijn Verburg is a Principal Software Engineering Group Manager at Microsoft, covering the Java, Go and Python ecosystems, and lives in New Zealand. His keynote was pitched as a warning shot for a room he suspected was mostly still using the chat interface.

“The genie has to be let out of the bottle before we can understand what the genie is going to do.”

• Adoption is off the charts and wildly uneven. ChatGPT alone has passed a billion active users, the fastest adoption of anything in human history, yet the people using fleets of agents to multiply their output number in the single-digit millions worldwide. He had 89 agents working for him while he spoke. Some of his engineers burn through ten to twenty thousand US dollars of tokens a week.
• These are no longer token-spitting machines. Mixtures of experts, agents that argue with each other before answering, deterministic tools for maths, chess and authoritative lookups. A colleague who is a professional mathematician said a proof that took two or three years now takes weeks; another threw his folder on the floor and quit after 35 years.
• Multimodal is disrupting creative industries first. His wife, a graphic designer, is going back to fine art because people now generate their own logos and decks. In defence, a former NZDF friend told him he would never sign up today.
• Software development has been democratised. He asked non-developers who had built something with AI to raise their hands; the room was full of them. At Creative HQ, where he mentors, startups that used to pay ten to fifty thousand dollars for a minimum viable product now build one over a weekend.
• Health and education are where he wants New Zealand to lean in. Radiologists are delighted to work through imaging with AI support and see more patients. His neurodiverse son learns through 45-minute nightly curiosity sessions where the AI asks questions back rather than handing over answers; he stressed the science of learning still says get the pen and paper out.
• New Zealand is last in the OECD on national AI strategy, in his reading. Ireland and Singapore have public, private and academic sectors pulling together on a national plan. He wants the same here, bipartisan, whoever wins the election, and quoted MBIE’s estimate of the GDP upside as reason enough.
• The risks are real. He described an agent swarm escaping a locked-down test environment and roaming the internet before anyone noticed, and made the case for regulation that keeps people safe while letting the private sector and academia move fast.

Greg Davidson: building the foundations for New Zealand’s AI future

Greg Davidson is Group CEO of Datacom. He delivered the densest thirty minutes of the two days, drawing on Datacom’s own adoption as “customer zero” and its 2026 State of AI Index, which Lou Compagnone, Datacom’s Director of AI, unpacked in more detail that afternoon. We’ve folded her session in here.

“An agent you can’t eval is just a demo.”

• Adoption is near universal; transformation went backwards. The index surveyed 207 senior leaders at New Zealand organisations with 100 or more staff. 91% use AI, up from 66% two years ago. 40% are exploring, 41% implementing, 15% scaling, and the share using AI to transform core operations halved, from 8% to 4%. Lou’s read: organisations aren’t going backwards; they have a clearer picture of what transformation looks like and are being more honest about where they are. 79% increased AI investment this year and a third say returns already exceed cost, but only one in five tracks the cost in a dedicated budget.
• It’s mostly chat. 68% use general-purpose assistants like ChatGPT, Claude and Gemini; 66% use AI built into productivity tools. That yields personal productivity, which vanishes into thousands of small improvements unless there’s a strategy to turn it into functional productivity: fewer handoffs, faster decisions, and work that no longer needs doing.
• The blockers aren’t technical. Lack of internal skills (20%), employee concern (18%) and implementation cost (17%) top the list, and 44% can’t identify a clear return or business case. Nobody said the technology doesn’t work. Lou’s point: the skills barrier dropped twelve points this year while employee concern rose seven, and the concern is about identity and control. People resist automating tasks bound up with who they are, and they stop trusting a system the moment they see it err, even if it’s more accurate than them on average. The fix for the second isn’t a better model; it’s giving people the ability to adjust or override. Training people to prompt doesn’t address either. Co-design does.
• Nobody owns it. 22% have a dedicated AI leadership role; many share responsibility across executives, which in practice means no one owns it. Only 39% have an AI strategy, which is why AI councils struggle to prioritise: they have nothing to prioritise against. Datacom’s separate Australian research on the chief AI officer role found 42% of organisations there already have one, largely because the federal government mandated one for every agency by July 2026. 74% of New Zealand respondents want a national AI framework and coordinating body; 64% want legislation on public-sector AI use.
• Two truths about agents. On OpenAI’s economic benchmark the best models beat a professional about 85% of the time when the job is to produce a deliverable. On the remote labour index, asked to finish a whole freelance project unsupervised, the best manage about 16%. Both are true; they answer different questions. Citing global enterprise surveys, Greg noted that most organisations have agent pilots but few have anything at production scale, and the top blocker is legacy integration.
• Evals are the unit tests of AI. You wouldn’t roll out C# written individually by every person for the same problem without testing it. Build a private test set, rerun it on every model release, and move agents through sandbox, shadow, human in the loop, human on the loop, then monitoring, one stage at a time and judged by evidence.
• Cheaper per token, dearer per task. Token prices have fallen more than 280-fold in eighteen months, but an agentic task uses around a thousand times the tokens of a chat, and the same task run twice can vary thirty-fold in consumption. Across Datacom’s customer base, cost per token fell 67% over fifteen months while tokens consumed rose more than 1,000%. Measure cost per successful outcome, cap retries and reasoning effort by default, and keep the CIO and CFO in the same room.
• Human in the loop is not enough on its own. Give an experienced radiologist a wrong AI reading and their accuracy falls from 82% to 45%. The Privacy Commissioner calls it automation blindness; Boeing’s MCAS system assumed pilots would notice and intervene within seconds, and 346 people died. A reviewer needs time, the same information the model saw, its confidence, its known failure modes, real authority to override, independent judgment recorded before seeing the machine’s answer, a path to dissent, and evidence that overrides happen. If nobody ever overrides, the review is ceremony.
• Know your digital supply chain. Four international cables carry 99% of New Zealand’s internet traffic, all landing in the upper North Island, and every cloud region in the country sits in one city. Six questions to ask of every cloud and SaaS service: where do my data, processing, identity, keys and logs sit; if the region is lost, is it a failover or a rebuild, and how many days; what upstream services does it depend on that I didn’t buy; who can compel access and under whose law; do I have a degraded mode; and when did I last test any of this rather than write it down.
• Open-weight models change the economics. Downloadable models now sit in the leading group on repeatable benchmarks, though the gap is widest on long-horizon agentic work, multimodal judgment and behaviour under attack, which is exactly what regulated deployment needs. Model choice is now a per-workload decision: GPU as a service for stable high-volume work, metered APIs for spiky frontier work. Downloadable weights don’t make you sovereign on their own; sovereignty is identity, keys, logs, licensing and the freedom to move.
• Where does New Zealand sit in the AI economy? He unpacked the value chain from chips and memory (Nvidia at roughly 60% operating margin, TSMC at 51%) through data centres, clouds, GPU clouds and frontier labs to the application layer. Every dollar in those boxes has to come back from the customer, and today’s discounted dollar is buying architecture lock-in tomorrow. The question isn’t whether AI is a bubble, but which layer you’re sitting in and whether the value you create outlasts the model you’re paying for.
• Six identities for the country, ranked by how much value stays here. Applied AI nation (agriculture, health, engineering, te reo); trusted operator running other people’s AI for sectors that can’t afford to get it wrong; enabler of sector experts who found companies; rule-maker middle power; compute exporter; and the one nobody chooses, passive customer that imports everything and decides nothing. His view is that we’re drifting towards the last one: 68% of New Zealand small businesses have no plans to invest in AI, against 38% in Australia. His recommendation is a trusted applied AI nation, with conditions on compute: every data centre brings its own generation, pays its own network costs, has anchored customers before consent, keeps ownership and operations here, and has mana whenua consent from day one.

His eight take-homes, four decisions and four drills: measure benefit, not usage; only evals turn on autonomy; measure cost per successful outcome; write the delegation contract before the agent goes live. Then map your digital supply chain; treat model choice as a workload decision; test your resilience in days and dollars; and plan hardware and AI consumption as a scarcity discipline, with twelve to eighteen months of lead time on critical kit.

Panel: sovereign capability in the AI era

Peter Griffin moderated Megan Tapsell (GM Enterprise and Pacific Technology, ANZ), Nick Valentine (Partner, Tech and Data, DLA Piper) and Martijn Verburg, picking up Greg’s challenge about which of the six identities New Zealand is drifting towards. We caught most of it.

• Australia is “night and day” ahead (Peter). Their AI strategy is being planned with a determination not to repeat the mining boom, where the upside flowed largely to foreign-owned companies and the country saw less of it than it should have.
• Sovereignty is not isolation (Megan). You can hold data in New Zealand and still have people offshore deciding how it’s used. What matters is the ability to participate in the governance and the decisions, for the country and for individuals whose data is being used to make decisions about them. The Māori-led conversation on data sovereignty has been running for years and the rest of the country can learn from it.
• Govern your dependencies, not just your models (Nick). We will depend on hyperscale capacity, offshore chips and frontier models, and that’s not failure. The sovereign layer is the middle: niche capability, sector expertise, trusted datasets in agriculture, health and mining. In contracts, that means exit rights, model portability, and the ability to leave if model behaviour, pricing or a provider’s geopolitical posture changes.
• What sovereignty a hyperscaler can offer (Martijn). Robust control over who can access data and why. No provider, hyperscaler or local, can claim everything is done in-house in an economy this size; Australia, with a similar rule of law, is a reasonable place for offshore resilience. He also noted that if Whakaari erupts or Wellington is flattened, critical data needs an offshore copy somewhere trusted.
• Regulation. New Zealand sits at the opposite end of the spectrum from the EU AI Act, relying on existing statutes. Nick thinks that’s the right call, but Australia and the UK have realised it leaves businesses to do the compliance mapping themselves and are centralising coordination. Copyright is the area where the light-touch approach is exposed: artists are finding their work in models without consent or benefit.
• Māori data sovereignty in practice. Datacom’s index found 20% of leaders say their organisation always considers Te Tiriti and Māori data sovereignty; 27% have never considered them.
• Small models and local compute. Martijn cited OpenRouter traffic data showing US companies increasingly routing to local or non-cloud-hosted models. You don’t need a frontier model to read a paragraph; multi-cloud and multi-model is a sensible way to keep prices down.
• The bank is not in a race (Megan). ANZ has slowed adoption right down under regulator scrutiny; the business is built on trust. Once you’re embedded with a vendor you’re embedded with their model and licensing, so cost control is hard. She also flagged an email that week from UBS requiring new hires to demonstrate AI competence, and expects that to spread.
• Who is “we”? Government for the national approach, business for investment and capability, and individuals who choose to be curious rather than remain consumers. Peter closed with a challenge to the room: politicians and policy people here know very little about AI compared with Australia, and it’s on the tech sector to help them.
• When a frontier model gets pulled. A question from the floor asked about the recent decision to restrict access to a frontier model to approved organisations, and what that says about our dependencies. Martijn’s answer: a US-headquartered platform company will comply with a direct order, and that’s true of any country hosting a model or a data centre. Having worked with the model behind locked doors, he thought giving hospitals, governments and emergency infrastructure time to patch before release was the right call. Very few of us need the absolute latest model to be successful.

Peter-Lucas Jones: small, community-driven AI at Te Hiku Media

Peter-Lucas Jones leads Te Hiku Media, one of 21 iwi radio stations, and was named to the TIME100 AI list in 2024 for its work on natural language processing for te reo Māori. His keynote was the most direct counterpoint of the two days to the “empires of AI” framing from Karen Hao’s book of the same name, which describes Te Hiku’s approach as the alternative.

“Trust’s friend is accountability, and its other friend is integrity.”

• Language loss is not an accident. Of the world’s roughly 7,000 languages, half are expected to be a memory by 2050. When a language goes, a way of seeing and categorising the world goes with it.
• Community-driven data. A ten-day reading competition drew more than 2,500 people who recorded Māori-language sentences, producing over 350 hours of tagged and labelled data. People opted in because they understood the purpose and trusted the station that had served their community for 35 years. Rongo, their pronunciation app, has been downloaded more than 50,000 times without advertising, and its users have contributed more than a thousand hours of speech data with consent given up front.
• Small models, one chip. Their first model was trained on a single GPU, not a data centre. Through Nvidia’s Inception programme they built a small training cluster in Kaitaia. The data is contained and knowable: 35 years of broadcasting, live video since 2014, interviews with kaumātua about every river, tree and marae in the region.
• A licence for the data. The Kaitiakitanga License protects the community’s data and sets the terms for who can partner with them. It allowed their Māori model to be fine-tuned for ʻŌlelo Hawaiʻi, a sister language, through a campaign that drew 1,400 Hawaiian speakers.
• What they offer. Under the Papa Reo platform: speech recognition for te reo Māori that they describe as the best in the world, New Zealand English recognition trained on voices people chose to offer, a bilingual model that transcribes both simultaneously, Kaituhi for transcription by researchers and broadcasters, on-device transcription for phones, text-to-speech for whānau who are non-speaking, live captioning, and Whare Kōrero, the app that brings all 21 iwi stations together. The platform is offered to other iwi stations and to businesses who want to collaborate.
• The challenge. Care about who governs, because these empires have more power than some countries, and concentrated power does not always mean a better future.

Human Rights Commission: a human rights and Te Tiriti approach to AI

Dr Stephen Rainbow, Chief Human Rights Commissioner, and Dayle Takitimu, the Commission’s Indigenous Rights Governance Partner, presented the report the Commission released in August, AI and Digital Technologies: A Human Rights and Te Tiriti o Waitangi Approach, built with an expert advisory group deliberately drawn from across the political spectrum, digital optimists and pessimists included.

• Only a third of New Zealanders trust AI, one of the lowest figures in the developed world. Human rights aren’t a barrier to innovation; they’re the foundation of innovation people can trust, and without trust the upside gets diminished.
• The frameworks already exist. International human rights law, domestic legislation and Te Tiriti provide the scaffolding; what’s missing is leadership and coordination. Our approach is light touch and fragmented, and that’s a big part of why we rank so poorly.
• Treat AI as infrastructure and govern it for the public good. The report’s core recommendations are a national digital infrastructure strategy, analogous to transport or water, and a national risk assessment that sets out a consistent, rights-protecting process for key decisions.
• Embed Te Tiriti from the beginning (Dayle). Māori are not waiting on the beach for salvation; iwi are often ahead of government at the AI frontier. The Commission’s advice is to build Te Tiriti and human rights into the design from the outset rather than taking the long road through the Waitangi Tribunal, the High Court and Parliament to be told, again, that it should have been there at the start. Māori at the table as designers and collaborators, not consumers.
• Three closing thoughts: if AI is infrastructure, govern it as infrastructure and ask who is included, who holds power and at whose cost; ground that governance in Te Tiriti and human rights; and create the conditions for innovation that is both transformative and trustworthy.

Panel: earning trust in the age of AI

Dr Mahsa McCauley, the AI Forum’s outgoing chair, moderated Dayle Takitimu, Emma MacDonald (Director, Centre for Data Ethics and Innovation, Stats NZ, and chair of the Forum’s social licence working group), Amy Dove (Forensic Technology Partner, Deloitte) and Professor Dr Aini Suzana Ariffin, who brought a UNESCO lens.

• Social licence is earned, not communicated into existence (Emma). Dig into the trust numbers and they’re specific: 39% don’t trust organisations with their data, 40% believe the technology is biased. People want to see the benefit for themselves, not just for big corporates. A useful rule: if you don’t want to talk about it, maybe you shouldn’t be doing it.
• Three obligations on organisations (Amy): the people affected have real influence over what gets built and whether it should be; governance is visible; and the responsibility is ongoing, because models and tools change.
• Consultation versus participation (Amy). If the decision to build the thing wouldn’t have changed with earlier community involvement, that was consultation. Participation starts before the decision to build.
• Principles are universal, implementation is national (Aini Suzana). UNESCO’s 2021 recommendation on AI ethics was adopted by 193 countries; more than 70 have since developed national AI policies and over 900 governance frameworks exist worldwide. UNESCO’s readiness assessment, covering governance, talent, infrastructure and data, science and education, and economics, has been run by more than 70 countries. New Zealand hasn’t done one yet.
• Accountability is simple (Emma). “You’re the adult in the room. You’re accountable.” No decision leaving your agency gets to blame the technology. UNESCO frames it as shared accountability across government, technology providers, network providers and end users, with guidelines for each.
• Data quality first. Small businesses in particular are struggling because they haven’t understood the importance of good data management before going anywhere near AI. Most of what Stats NZ does with AI isn’t generative; it’s machine learning in statistical methods.
• Would ten-times more accurate models fix trust? No. People already use tools they say they don’t trust. Social licence bites when something goes wrong. Nobody wants to be the agency that delivers New Zealand’s robodebt or the Netherlands’ welfare scandal, and from the public’s side “no one cares which department I’m from, I’m just from the government”.
• Trust isn’t permanent (Aini Suzana). What earns it today may not tomorrow; continuous assurance is the job.

Panel: from pilot to production

Stacey Morrison moderated Stacy Pence (AI and Data Lead, Accenture ANZ), Yash Kapoor (CEO and founder, Innovate Now) and Jayne Foster (Principal Advisor, DPMC). We joined partway through.

• Four things the successful ones have (Stacy). A leadership team bought in top to bottom; an architectural framework, even a loose one, so the technology team knows where to point; a learning culture where people get excited rather than scared; and a plan for workforce transformation. Most organisations have two of the four.
• Start with the problem, not the tool (Jayne). DPMC has worked with agencies on common policy problems such as duplicated effort and large volumes of submissions, using Creative HQ to run a design process first. Government has to be a fast follower rather than frontier; there’s more to consider, including Māori data sovereignty, and a fine line between enabling innovation and smothering it from the centre.
• Fix the foundations before you layer AI on top (Yash). If your data isn’t ready, putting another layer over it doesn’t solve the problem. Domain experts at the front line understand what fails and why; leadership’s job is to give them and the technologists a common conversation.
• Think in outcomes, not tasks (Stacy). Bank customer onboarding is twelve to fifteen steps and 48 hours because humans check things in sequence. An agent that collects what’s needed and goes back to the customer in real time can collapse that. Humans think linearly; agents don’t have to.
• Realistic timeframes (Yash). Treat it like hiring someone: a 30, 60, 90-day cycle, and trim the scope to what delivers value in the quarter. Small wins build the trust for bigger ones.
• Pilot data versus production data (Stacy). Clean up what you’re letting the AI touch, define it, and decide how far you’re taking it; an agent drafting an email for review needs a different data readiness from one that sends it.
• Priorities for the next twelve months. Jayne: policies, guidance and upskilling, so the organisation is ready to move. Stacy: pick one business function and go deep, then use the proof point to scale. Yash: decide whether you want AI in a process or an AI process (the second needs reimagining), have a realignment plan for your people, and rethink how the operating model measures value.

Kiwibank: rebuilding a bank with AI in the delivery lifecycle

Romain Groleau and Stacy Pence from Accenture interviewed James Woodward, Kiwibank’s transformation director. Kiwibank is about three years into rebuilding the bank: core banking, lending, data, cards and payments, cyber and customer channels, with an ambition to double its customer base and financial performance by 2030.

• Where AI shows up most is in delivery. Discovery, analysis, development and testing, using knowledge graphs and content layers to accelerate. Agents built for the analysis stage produce clearer requirements and user stories, which feeds faster and higher-quality technical delivery downstream. Accelerate the left of the lifecycle and the whole lifecycle accelerates.
• Principles first. Value-led (no AI where there’s no strong baseline to measure against); AI as an amplifier of people’s experience and judgment; 100% human in the loop at this stage; and platform-led, because platforms bring the guardrails that make the risk manageable.
• Culture from the top. The CEO sponsors the AI programme personally. They’ve invested in the literacy of leaders, many of whom have to learn to lead in a world where AI is a much bigger part of the job. Copilot across the organisation delivered small individual gains but a huge playground for people to learn what AI is good and bad at.
• Stop forcing the journey. James admitted to trying to push people along early on. Now they let people come at their own pace with encouragement, because forced adoption just breeds the scepticism that undermines it.

TVNZ: bringing tikanga to the front of an AI system

Tuscany Joelle-Hui, Portfolio Product Manager for Data and AI at TVNZ, and broadcaster Stacey Morrison showed how TVNZ is building agentic metadata generation for its video archive while honouring its te ao Māori commitments.

• The opportunity is understanding video at scale, not generating it. Agent teams create metadata that helps people triage content, cut promotions and highlights (currently very manual), and place advertising in the right moment. A responsible archive is the next piece, including maintaining tino rangatiratanga over Māori content.
• Tikanga at the front, not as a checkbox. A Kaimanaaki agent greets the user at the start of the experience, explains why it’s there, and prompts the questions that out-of-the-box tools can’t: does this look like Aotearoa, who is missing, how is this whānau or this taonga represented, and whose mana is attached to it.
• Three principles. Lead with partnership from day one, even when it’s uncomfortable. Deliver real proofs of concept in a safe environment, technically (with AWS, where the content already lives) and in terms of which content and talent are involved. Don’t retrofit: you can’t build an AI system for Aotearoa without tikanga, and if you’re adding it afterwards you’re just adding governance and missing the opportunity.
• Don’t let perfect be the enemy of good. Nobody is the perfect person for this job; you need the right connections and a willingness to be vulnerable and make mistakes in a low-risk space.

Hackathon finalists

Four teams from the Aotearoa AI Hackathon pitched what they had built in a day and a half.

• A marketplace for primary-sector waste. Producers describe a waste stream in plain language; agents profile it, match it to local reuse or recycling, and if nothing fits, link it to known research on future uses, building a national map of untapped by-products along the way.
• Te Ara Taea, a field platform for Individual Placement and Support (IPS), from the Wise Group, a mental health NGO. IPS doubles the odds of employment for people with mental health and addiction challenges, but providers were finding employers via Google and driving down the street. The platform handles caseloads, employer discovery (including businesses that never advertise), visit planning with route optimisation, and voice-to-record notes in ninety seconds. Tested with the teams who would use it.
• Kai Line, grocery shopping by phone call for the 180,000 New Zealanders who are blind, deafblind or have low vision, and the people who care for them: an agent that takes a natural conversation in more than 100 languages, builds the order, checks specifics like apple variety, and arranges delivery.
• Kai Rescue, a matching and routing engine for New Zealand’s food-rescue network. Donors photograph surplus food, AI reads weight, allergens and freshness with a human checking, recipients describe what they can handle, and an optimiser matches and routes drivers. Their pitch: New Zealand doesn’t have a food shortage problem, it has a food routing problem.

Congratulations to the Wise Group, who took the overall win, and to Kai Line, who won the new Technical Brilliance award by a unanimous judges’ vote.

David Steele: AI as a team sport

David Steele directs Plug and Play Tech Center’s operation across Texas. He’s a coach, not an AI expert, and framed the whole talk around a high-school football play he got wrong, then got right once he listened to his coach and trusted his teammate.

• Nobody wins this alone. Four players have to be at the table: startups (the disruptors), corporations (chasing growth and hampered by their own processes), cities and governments (setting the guardrails and the incentives), and universities (where research sits on a shelf until a founder commercialises it).
• How Texas did it. Five cities each picked a lane: sports tech in Frisco, AI in McKinney, life sciences and smart cities in Sugar Land, aerospace and advanced manufacturing in Cedar Park, and Bryan alongside Texas A&M. Eighteen companies relocated in three years because the right people were in the conversation from day one.
• Four plays for New Zealand. Know your position: which of the four players are you. Pick one problem, not every problem (“you’ve got enough geothermal activity to boil the ocean already”). Set one table with all four players at it. Then rep it: practice, practice, practice.

“Three years ago, if you weren’t using AI, you were behind. Today, if you’re not building with it, you’re behind. AI is not replacing jobs. People who know how to use it are.”

Day two opening: what future are we building?

Maria Mingallon, the AI Forum’s incoming chair, opened day two with a recap and a personal frame. The risk that worries her most is not that AI becomes too powerful but that too many people are left behind, and in particular the graduates and interns whose entry-level work is being automated. Her argument: we never hired graduates for the mundane tasks; we hired them to challenge how we work, and that is exactly what a successful AI transformation needs. They deserve better opportunities, not fewer. She quoted the Microsoft and Accenture estimate that AI could add $76 billion to New Zealand’s GDP by 2038, and closed with a request: look into the eyes of a child for thirty seconds and ask what future they will inherit.

Accenture: workforce transformation

Louise Barrere, who leads Accenture’s global relationship with OpenAI, and Dr Viveca Pavon-Harr, global AI and data lead for its health and public sector practice, kept the human at the centre, starting with Louise’s twelve-year-old son and the hearing aid he has worn since birth. His one request to the manufacturer was a charger so he never has to leave class or stop what he’s doing to change a battery. The new device, released this year, has it.

• Ask what breaks the flow of work. That’s the question to take to your team: which tabs, handoffs and waits pull them out of what they’re doing. It’s a better starting point for AI than a technology roadmap.
• Stop documenting the current state. A year or two ago the method was as-is, to-be, incremental improvement. Now: a one-page future-state vision, the jobs to be done and new roles, the data that powers them, a build-versus-buy decision, and a roadmap with change and adoption. Benchmark the current state for value tracking, then move on.
• Just because you can build it doesn’t mean you should. A client recently dropped a long-standing software provider to build the capability itself with AI. Now they own the maintenance, the innovation roadmap and the security of that product forever. Be clear about total cost before you replace a vendor with a demo.
• Henry Ford’s thirty years. Electricity arrived in factories around 1890. Ford’s first move was to drop electric motors into a plant laid out for steam; the gains were marginal. When he built a new plant designed around electricity and the moving line, the build time for a Model T fell from over twelve hours to about 93 minutes and the price dropped around 70%, but it was the 1920s before that flowed through. We don’t have thirty years.
• Tokenomics in practice. A clinical-trials delivery pod went from eight people to four and consumed two billion tokens doing it. The reaction wasn’t to shut it down; it was to ask whether the output was valuable, then optimise and make the workflow reusable across every other pod. Expect the bumps, then build the reusability.
• Governance as a lanyard. Everyone at the summit wore one: a visible signal that you’re registered, legitimate and here for a reason. AI systems should carry the same visible certification of what’s running, on which data, under whose authority. The black box shouldn’t be.
• Pilots are over. In the last three months their public-sector work has shifted from demos and pilots to “phase one”, designed to scale with heavy testing up front, because the client burnout on flashy demos is real. Benchmark models on your own processes rather than external leaderboards; you rarely need the top model for the task.

Panel: Augmenting Human Potential

Louise Barrere moderated Ana Ivanovic-Tongue (Chief Delivery Officer, academyEX, and chair of the Forum’s AI Skills Working Group), Nilay Rathod (GM Group Architecture, Spark), Amanda Veldman (Senior Learning and Development Advisor, Datacom) and Te Rohu Crowe, a final-year law and arts student at Victoria University of Wellington.

• The entry-level pipeline is a leadership decision (Ana). The media landscape is polarised between “only 5% affected” and “half of us lose our jobs”, and the truth is we don’t have the data yet. How you grow the pipeline of your next leaders isn’t a question of which tasks are automatable. New Zealand is heading from four working people per retiree towards two within fifteen years; AI isn’t optional for the country either.
• The graduate view (Te Rohu). AI screening cuts the entry-level pool and removes the human connection; three rounds of talking to a computer before meeting a person. Among students, opinion is sharply divided between those who won’t touch it and those who use it for everything, and the gap is mostly education about how to use it as a tool.
• Intentionally create the opportunities (Amanda). If AI takes the tasks trainees used to learn on, organisations have to deliberately build new ways in. It isn’t a luxury; it’s a responsibility.
• What Spark has done (Nilay). Copilot for test automation lifted productivity 30 to 50%, but the real change is that the team now spends its time on security and quality rather than writing test scaffolding. Call summarisation freed contact-centre staff from post-call notes. AI is changing the workflow itself; the education people need is on how the future workflow looks, not just the tools.
• Career capital, not tool skills (Ana). Domain capital (the tacit knowledge of your sector), AI and data capital (including probing for accuracy and knowing your own biases as well as the model’s), judgment capital (when to use AI and when not, and the accountability that comes with it), relational capital (collaboration and holding judgments lightly) and adaptive capital (relearning as a daily, communal exercise).
• Tool-agnostic microlearning (Amanda). Tools change weekly; she can’t update a hundred modules to keep up. Teach transferable skills in fifteen-minute chunks people can apply immediately, and have leaders model curiosity, including sharing what they tried that didn’t work.
• A weekly hour on the AI software lifecycle (Nilay). Spark runs an open forum every week on how AI is changing design, build and test, is upfront that the benchmarks are heading towards near-perfect scores on software tasks, and commits to keeping people future-fit whether they stay or go. Some engineers who love writing code have opted out, and that’s been treated as legitimate feedback.
• Archetypes (Ana). She pointed to Boris Cherny’s five archetypes for how people work with Claude Code: Prototyper, Builder, Sweeper, Grower and Maintainer. Understanding people’s natural inclinations, then augmenting around them, is a better conversation than “everyone learn the tool”.
• Inclusivity (Amanda). Equitable access does not produce equitable outcomes. People need to understand how these tools fail, where bias comes from, and feel able to push back, or the bias flows downstream. Nilay: get the right voices in the governance group (people and culture, risk, legal, security), and put ownership of each agent with the business function that uses it, not with technology. Ana: within twenty to twenty-five years half the working population will be Māori, Pasifika or migrant, and the leadership tables deciding what values AI upholds at work don’t yet reflect that.

John-Daniel Trask: the AI opportunity for New Zealand

JD Trask co-founded Raygun, which is 93% export revenue, and more recently Autohive, an agent platform that grew out of the AI systems Raygun built to run itself. His keynote was the most provocative of the two days, and he knew it.

“Three and a half years in and we’re still having 2023 conversations.”

• Sell more to the world than we buy. Software removes the tyranny of distance: milliseconds to market instead of months for a container of planks. AI lets New Zealand double down on that. Our two biggest export industries, food and tourism, are largely immune to AI disruption and may get tailwinds from it; tech is the one to push.
• What we’re missing is ambition, leadership and appetite for risk. In his view the country has spent three and a half years talking about governance, ethics and burnout while almost nobody builds. At Raygun those topics were “a bullet point in a leadership meeting”, after which they got on with it.
• Don’t build a model. If New Zealand wants sovereign AI, build data centres, fine-tune open-source models on local concerns, and make them available to everyone here. He’d raised the risk of export controls on intelligence with a political party months before the recent restriction of a frontier model landed.
• Small is now a superpower. Microsoft, Amazon and Meta are in quarterly layoff loops because they know they don’t need 250,000 people any more, and the layoffs destroy the culture that innovates. New Zealand’s small businesses don’t have that problem. The question isn’t who to cut; it’s what would this business do with 50% more headcount, then get AI to do that part.
• Chief AI officer is an anti-pattern. Does yours have lunch with the chief electricity officer? AI is an “everybody technology” and has to be led from the top, by people willing to make the same mistakes as their teams and put money behind it. He paused Raygun for a week in 2023 to have everyone build agents.
• Motion is not progress. Replace steering committees with thirty-day return-on-investment pilots, owned by a C-level sponsor paired with a hands-on builder, with money as the metric. Scale what works, kill what doesn’t, run the next one.
• The goal state. More work happening while the team is asleep than while they’re at work. One of their products had its cost to serve cut 95% overnight by an agent while the engineers slept. As output multiplies, shorten the working day; that’s what employees should get back.
• His advice to individuals. If your employer isn’t investing in AI, leave before you’re pushed. Someone whose experience is a couple of prompts into Copilot wouldn’t make it to an interview at his companies.

Closing

Maria Mingallon closed the summit on three threads: human potential first, with productivity measured in higher-value work rather than time saved; stay curious, and keep asking why we do it this way; and New Zealand’s smallness as a competitive advantage if we choose to take it. Her admission, as someone working inside a large corporate: AI-native consultancies will be able to do what the big firms do with fewer people and in ways we can’t yet imagine.