The change is on the scale of the climate transition — and unlike climate, we are still early enough to choose our path. Draft for discussion, June 2026.
Status: draft for discussion. This document is offered to every party in Parliament, to officials, and to the public, as a contribution to a decision that belongs to all of us. It is not a party document and it is not a product. It is expected to be marked up and improved.
In June 2026, a single overseas order disrupted access to a major AI tool for users worldwide. New Zealand’s public services and economy increasingly run on tools like it — built, hosted, and controlled offshore. The ability to change them, to mine what we feed them, or to switch them off sits with others. We are giving that control away by default, one contract at a time, without ever deciding to.
This is not a contest between two silos, and it is not about who leads this year. Today the United States stack draws the most adoption and goodwill, and the single most capable model in mid-2026 is still American. But that lead is neither permanent nor the point. The strongest Chinese open-weight models — Alibaba’s Qwen, DeepSeek and Zhipu’s GLM — now match or lead the best open American ones on public benchmarks and dominate the open-source developer ecosystem, and they do it far more cheaply: Qwen is the most-derived model family on the main open-model platform, and named US companies already run these models in production — Airbnb’s chief executive says it relies on Qwen for customer service because it is “fast and cheap”. Among large enterprises in production this is still early rather than mainstream, but the driver is plain. Open-weight models built offshore can be run several times more cheaply per token than the leading US services, and many times cheaper again at the lower tiers, so cost is steadily pulling workloads towards whichever stack is cheapest to operate. Depending on a Chinese stack would carry the same exposure as depending on an American one — the control, the off-switch, and the access to what we feed it would simply sit in a different capital. The decision New Zealand faces is long-term and supplier-agnostic: not which foreign supplier to trust, but how to keep custody and control at home, whoever leads at the time.
Artificial intelligence is reshaping how New Zealanders work, how our public services treat us, and who decides the matters that affect our lives. Changes that large, and that lasting, are not well settled by one government and reversed by the next: they call for policy, and for a direction the whole Parliament can hold in common. Treated as a partisan contest, this will be captured, watered down, or left to drift into choices made elsewhere. Yet the direction itself is hard to argue against, because at heart it is not about technology. It is about rights New Zealanders already hold, carried into the age of AI: to control what is done with your information, to a human decision on what affects your life and your livelihood, and for a people to govern what is theirs. No party campaigns against those. And we are unusually well placed to lead — small enough to move quickly, trusted enough to be heard, and early enough that the path is still open. New Zealand can help shape the path others take, rather than inherit one it had no hand in. The question is whether we choose to.
For busy MPs: a one-page summary is available — read it or download the PDF.
Artificial intelligence is moving into New Zealand’s work, public services and everyday decisions faster than our law and policy are keeping up. A useful comparison is the climate transition, like electrification or the arrival of the internet: a structural shift that touches every sector, rewards countries that decide early, and penalises those that drift. The difference, and the opportunity, is that on AI we are still early. The infrastructure is only now being laid, the rules are still being written, and the habits of use are still forming. New Zealand can still choose its path rather than inherit one made elsewhere.
This proposal sets out a direction a majority of New Zealanders can recognise as their own, and that parties across the House can stand on together without surrendering what makes each of them distinct.
It rests on a single, achievable idea of sovereignty: not owning the chips, data centres and frontier models — a race a country our size cannot win and need not run — but authority, custody and control over our records and our decisions. That foundation is within reach, and everything else is built on it.
The case is urgent because two exposures are already live and growing. The infrastructure we depend on is foreign-owned and reachable under foreign law, and access to it can be switched off without our consent — as a single United States order demonstrated in June 2026. And every day, sensitive New Zealand information flows into foreign AI tools through ordinary use: in the largest cross-country study, 48% of employees had put sensitive company or customer data into public AI, and 56% had used it at work without knowing whether they were allowed to.
The proposal offers a platform of seven common-ground commitments — authority stays here; people move up, not out; sensitive information stays home; we respect Te Tiriti and Māori data sovereignty; people decide, machines assist; what public AI does is checkable; and we build our own capability, with others — and a three-phase pathway to deliver them:
Almost all of this repurposes machinery New Zealand already has. Only one genuinely new function is asked for: an oversight-and-register body for public-sector AI.
The ask: convene a cross-party, multi-stakeholder process to set this direction — the concrete first step being a select committee inquiry — and stand together on the seven commitments. The window will not stay open. The choice is not whether AI arrives — it already has — but whether it runs on our terms or someone else’s, while that is still ours to decide.
Every generation faces one or two changes large enough to reset the terms of national life. For ours, artificial intelligence is shaping up to be one of them. It is moving into the work people do, the services they rely on, and the decisions made about them, and it is doing so faster than our law and policy are keeping up.
The closest comparison we have is the climate transition. Like climate, AI is a structural shift that will touch every sector, every region and every household. Like climate, it rewards countries that decide early what they want and act while the path is still open, and it penalises those that drift and inherit choices made elsewhere. The difference, and the opportunity, is that with AI we are earlier. The infrastructure is still being laid, the rules are still being written, and the habits of use are still forming. New Zealand can still choose its path rather than have one handed to it.
This proposal is written for that choice. It is offered to every party in Parliament, to officials, and to the wider public, with a single aim: to set out a direction a majority of New Zealanders can recognise as their own, and that parties across the House can stand on together without giving up what makes each of them distinct. It is not a party document and it is not a product. The questions it raises are too large, and too long-lived, to be settled by one government and undone by the next. They are the kind of questions a country is better off answering as one country.
The argument is straightforward. AI used well can lift people: take the drudgery, leave the judgement, and move people up rather than out. AI used carelessly can hollow out work, move decisions that should stay human into systems no one here can see into, and quietly shift control of our information offshore. The deciding factor is not the technology. It is whether New Zealanders keep authority over the AI we depend on, and over the records and decisions that AI touches.
The usual debate about AI asks what the technology can do. That is the wrong starting point for policy. The questions that matter to people are simpler and more human: which of our work, and which of our decisions, should stay in human hands, and do New Zealanders keep custody of them.
Three things are at stake.
The first is the character of work. AI can be used to lift people into better, more skilled and more interesting work, or to remove them from it. Both are possible with the same tools. Policy that measures success by adoption figures alone will not see the difference. Policy should be judged by what it does to people and communities, not by how much AI gets bought.
The second is control over decisions that affect people’s lives. As AI is built into public services and private systems, decisions about entitlements, risk, eligibility and opportunity can drift into automated processes. Some of that is useful and safe. Some of it is not. A decision that materially affects a person should rest with an accountable human, and where a question cannot be reduced to a clear rule it should return to human judgement rather than be settled silently by a machine.
The third is custody of our information. New Zealanders’ personal information, and the commercial information of New Zealand businesses, is increasingly fed into AI systems that run offshore, under other countries’ laws. The value is real, but so is the exposure. Information that leaves the country, and leaves our legal protection, is no longer fully ours.
Sovereignty, in this proposal, means something concrete and achievable. It is authority, custody and control over our records and our decisions: the ability to know where our information is, who can reach it, and on whose terms. It does not mean owning the chips, the data centres or the frontier models. That is a race a country our size cannot win and does not need to run. The achievable foundation is control over what is ours, and that is within reach.
New Zealand has made a start, and it is a sensible one as far as it goes.
In July 2025 the Government released the country’s first national AI strategy, Investing with Confidence (MBIE). Its focus is adoption: encourage wider use of AI, rely on voluntary guidance, and add no new law. Adoption matters, and lifting New Zealand’s use of AI is a reasonable first priority. But the strategy is silent on the harder questions: who controls the AI we depend on, where it runs, under whose law, and where our information goes when we use it. It says little about procurement, data residency or enforceable oversight.
The other pieces in place are also voluntary. The Public Service AI Framework (2025) is non-binding and maintains no public register of the AI government uses. The Algorithm Charter for Aotearoa New Zealand (2020) is a voluntary commitment by Crown agencies. None of these creates an obligation that can be checked or enforced.
Civil society has moved into the gap. The AI Forum’s AI Blueprint for Aotearoa (2025), the body of Māori data-governance work, and other contributions have begun to set out what government policy has not yet provided. New Zealand is not short of thinking. It is short of a settled, cross-party direction with enforceable commitments behind it.
This is the opening this proposal addresses. The first national strategy chose adoption and left control for later. The argument here is that control cannot safely be left for much later, because the exposure is already running, and growing, while we wait.
A note on te Tiriti runs through everything that follows: Māori data is a taonga, and decisions about it sit with Māori-led frameworks. This proposal does not speak for Māori or displace that work; it is built to be implemented alongside it.
Two exposures are already live. Neither is hypothetical, and both are getting worse the longer they go unaddressed. One is in the infrastructure we depend on. The other is in what we send through it every day.
New Zealand has almost no domestic frontier compute. Our first hyperscale cloud region opened only in December 2024, and it, like those expected to follow, is United States-owned. Many public agencies use Microsoft cloud services under the all-of-government agreement, historically hosted from data centres in Australia, and the Government’s long-standing “Cloud First” policy actively directs agencies towards cloud services. The practical effect is that a large share of New Zealand’s public and private data sits on infrastructure owned and operated by foreign companies.
That matters because data held on these systems is reachable under foreign law regardless of where the servers physically sit. The US CLOUD Act (2018) compels American providers to hand over data they control, wherever in the world it is stored. FISA Section 702 authorises surveillance of non-US persons abroad through US providers. China’s National Intelligence Law (2017) compels Chinese firms to assist the state. These are not edge cases. European data-protection authorities (the EDPB) have found the CLOUD Act in conflict with their privacy law, and before the French Senate in 2025 Microsoft France confirmed it could not guarantee that French public data would be shielded from US authorities.
The reach extends to access itself, not only to data. On 12 June 2026 the US Commerce Department issued an export-control directive requiring Anthropic to suspend access to two newly launched Claude models for any foreign national worldwide, including foreign nationals inside the US and the company’s own staff. Because the company could not reliably identify its users by nationality, the practical result was that it initially disabled both models for all users globally (reported by Al Jazeera and Nextgov, 13 June 2026; access was later restored to a set of vetted partners; the legal designation is contested by some analysts, so it is described here as an export-control directive). Whatever its precise legal basis, the episode showed the pattern plainly: access to capability we might come to rely on can be switched off by a foreign authority, with no New Zealand say in the decision.
For a small trading nation, the decisive risk is not a single dramatic breach. It is that the control, the off-switch, and the everyday flow of our information all sit offshore, exercisable without our consent.
The second exposure is closer to home, and most organisations do not see it. It is not in the infrastructure; it is in what people pour through it.
Every time a staff member pastes a document, a customer record or a strategy note into a public AI tool, that information leaves the organisation. On the consumer tiers most people use, inputs are, by default, retained and used to train future models unless the user opts out. The prompt is the leak.
The scale is now measurable. The most comprehensive cross-country study to date, by Melbourne Business School with KPMG (2025, drawing on a 32,352-employee subset of a 48,340-person survey across 47 countries), found that 48% of workers had entered sensitive company or customer information into public AI tools, and 56% had used AI at work without knowing whether they were allowed to. A 2024 survey by CybSafe and the National Cybersecurity Alliance, which included New Zealand, found 38% of workers share sensitive work information with AI without their employer’s knowledge.
This is happening now, across the economy, largely unsanctioned and largely invisible to the people accountable for the information. An adoption-first message that says nothing about it only speeds it up. The answer is not to switch AI off, which surrenders the benefit. The answer is to give people AI they can use safely, kept within New Zealand’s control and legal reach, so the usefulness stays and the information does not walk out the door.
The exposure is not spread evenly. Three areas carry the most weight, and the ranking that follows is an analytical judgement about where the dependency bites hardest, not a measured league table.
The public sector and the courts are the most directly exposed, with more than 180 agencies already running on offshore infrastructure and handling information that citizens have no choice but to provide.
Health and finance follow. Both are heavily regulated and hold some of the most sensitive information about New Zealanders, and both increasingly run on the same foreign stacks. A control or access decision made offshore reaches straight into regulated, high-sensitivity systems.
Small and medium businesses and the primary industries are the least equipped to protect themselves. They have the thinnest in-house IT, the least capacity to re-platform, and the least ability to recover if a model or service they have come to depend on is withdrawn. They are also the backbone of the economy. If the response to AI leaves them out, it leaves out most of New Zealand.
These are the people and institutions a cross-party direction needs to protect. The sections that follow set out what that direction can be.
When vendors talk about “sovereign AI”, they usually mean owning the whole stack: the chips, the data centres, the frontier models. That is a race a country our size cannot win, and need not run. The capital is enormous, the technology turns over every few years, and the leading edge sits with a handful of firms in much larger economies. We should set that meaning aside at the outset, because chasing it would waste money we do not have on a contest already lost.
Real authority over AI starts lower down, and it is within reach. It begins with a plain, verifiable property: that your records, and the proof of what was done with them, are yours. Held by cryptography and by possession, not granted as a permission an operator can withdraw. This is the record-foundation. Get it right and the larger claims become real and enforceable: authority over your data, over how a system acts for you, over what the state’s AI may do in your name. Without it, every grander claim of sovereignty is only a word. We use the term “sovereignty” sparingly in this proposal, and we mean it in this sense alone: authority, custody and control over records and decisions, not ownership of the machinery.
Three plain tests follow from the record-foundation. They are deliberately simple, so a board, a council or a select committee can apply them without a technical adviser in the room. They apply equally to what government buys and to how the rest of us use AI.
Hold and verify, rather than trust. A buyer should be able to inspect and check what a system actually does, not rely on a distant vendor’s assurance that it behaves. Trust that cannot be checked is not a safeguard; it is a hope. Where the record of what a system did is held in your own possession and signed so it cannot be quietly altered, you can verify rather than take on faith.
Build the protection in. Safeguards should be enforced by how a system is engineered and procured, not asserted in a voluntary principle that no one audits. A rule that depends on a supplier’s continued goodwill fails at the moment goodwill is withdrawn. Where protection rests on cryptography and possession rather than on good behaviour, it holds even when the commercial relationship turns or the supplier is compelled by a foreign authority.
Fit the context, not the average. A model scoped to a specific community, sector or task, accountable to people here, will usually serve better than one global-average model accountable to no one in this country. Local fit is often the more accurate and more governable choice, and it keeps the people who understand the context in the loop, rather than deferring to a system trained on everywhere and nowhere.
These three tests also answer the everyday leak, the steady flow of sensitive information into foreign tools that the adoption-first approach leaves unaddressed. The response is not to switch AI off. That is abstention, and it surrenders the benefit while competitors keep it. The answer is custody: AI kept under your own control and legal reach, so people get the usefulness without the information leaving the country or its legal protection. Custody is the practical form the record-foundation takes in daily work. It is what lets a clinician, an accountant or an official use these tools on real cases without the case walking out the door.
Underneath the technical detail there is a short list of commitments that few New Zealanders, and few parties, would reject. They are not left or right. They are the terms on which a small democratic country keeps hold of its own future, and they are offered as a platform that parties across the House can stand on together without surrendering what makes each of them distinct.
1. Authority stays here. New Zealanders keep rightful authority over the AI that runs our public services and holds our data. No party campaigns to hand control of the public’s records, or the off-switch to public systems, to a body beyond New Zealand law. Framed as national control of national infrastructure, this reads as common sense across the spectrum.
2. People move up, not out. AI should be used to lift people into better work, taking the drudgery and leaving the judgement, not simply to remove them. A party of enterprise can back the productivity gain; a party of workers can back the protection of livelihoods; both can agree the aim is people in better jobs rather than fewer jobs.
3. Sensitive information stays home. Information about New Zealanders and our businesses should not have to leave the country, or our legal protection, to be useful. This is a security and a commercial point before it is anything else, and it speaks equally to protecting citizens’ privacy and to protecting firms’ commercial advantage.
4. We respect Te Tiriti and Māori data sovereignty. Māori data is a taonga, and te Tiriti obligations are met in how public AI is built and bought. This is grounded in te Tiriti and in Māori-led instruments already in place, so it builds on established settings rather than reopening them. The detail is led by the Māori-led frameworks that govern this ground; this proposal honours that lead and does not speak for Māori.
5. People decide, machines assist. Decisions that materially affect people stay with accountable humans, and where a question cannot be reduced to a rule it returns to human judgement. Every party wants a named person answerable for a decision that affects a citizen, rather than an unaccountable system. The principle protects both the public, who keep someone to appeal to, and the elected, who keep the authority they were chosen to exercise.
6. What public AI does is checkable. Government AI keeps an auditable record of what it decided and why. Transparency in the spending and decisions of the state is a value every party already claims, so this only extends an accepted principle to a new tool.
7. We build our own capability, with others. New Zealand develops the capacity to do this itself and in partnership, rather than depending on a switch held offshore. This frames as backing local skills, local firms and national resilience, which crosses every party’s economic story. “With others” keeps it open rather than protectionist, so it reassures those wary of going it alone while still answering those who want capability kept at home.
Taken together, these seven are the substance a cross-party process could adopt without anyone giving up their distinct programme. They set the direction; the parties keep their own routes to it.
New Zealand is not the first small advanced economy to face this. Several comparable countries have already built instruments we could adapt, and the value of looking at them is practical: it shows the moves are tested, affordable at our scale, and need not be invented from scratch. The point is to transfer the instrument, not to copy a larger country wholesale.
The European Union has set a market-access benchmark through its AI Act, which triages systems by risk (unacceptable, high, limited, minimal), requires transparency labelling so people know when they are dealing with a chatbot or a deepfake, and places a duty on organisations to build AI literacy. For an exporter, interoperability with the EU is an asset worth keeping in view: aligning where it eases market access, without adopting the full regime, is the sensible posture for a country our size.
Canada treats public investment as a lever. It has set up a public compute-access fund for small firms and researchers (the AI Compute Access Fund, up to CA$300 million within a larger sovereign-compute strategy), and it maintains an “AI Source List” of pre-approved suppliers so agencies can buy with confidence. Government also acts as a strategic anchor customer, using its own demand to pull domestic capability into being. The supplier-list idea maps directly onto the pre-vetted procurement path in this proposal.
Singapore pairs money with hand-holding. Its Enterprise Compute Initiative offers cloud credits plus consultancy worth up to S$150 million to help firms build a working first version, and a separate National AI Impact Programme aims at the practical end, with targets of around 10,000 enterprises supported and roughly 100,000 workers trained by 2029. The lesson is that credits alone are not enough; the advisory support is what gets smaller organisations over the line.
Ireland shows the lighter-touch path: a refreshed National AI Strategy in 2024, an SME-awareness campaign run through Enterprise Ireland, and skills delivered through the existing Springboard+ programme. This is capability-building routed through machinery a country already has, which is close to the approach this proposal favours.
Estonia demonstrates capability in the public sector itself. Its Bürokratt conversational gateway gives citizens a single way into government services and reportedly spans more than 200 use cases, supported by a national AI-skills programme. It is evidence that a small state can build genuinely useful public AI rather than only regulating others’.
Denmark is the closest to the record-foundation argument here. It has built a sovereign public-sector AI capability running on home and EU-jurisdiction compute, so that inference and data sit under home law, and paired it with a regulatory sandbox for safe experimentation. This is custody as policy: keeping the processing, not just the storage, within reach of domestic law.
The cross-cutting lesson is the one most relevant to us. Denmark and Canada both treat home or EU-jurisdiction compute as the decisive lever, because keeping inference and data under home law is what turns a paper commitment into a real one. That is precisely the record-foundation reframe in section 5, already in practice in countries of comparable size. None of these instruments requires owning the frontier; each is affordable, tested and adaptable to the New Zealand context.
The work described in this proposal is one connected sequence, not a set of separate programmes. It runs in three phases. Each has a realistic New Zealand owner and uses a lever that already exists wherever one is available. The order is deliberate: set the rules first so that public demand pulls domestic capability into being, then fund the capability that meets that demand, then govern and sustain it. Indicatively, Phase one begins within the first year, Phase two is established across years one to three, and Phase three’s institutional arrangements are put in place alongside.
These moves cost little and can begin immediately, because they change rules and standards rather than build infrastructure.
A procurement test on jurisdiction of inference and data residency. This is a pair of tests added to public-sector AI buying: jurisdiction of inference (where a model actually runs and whose law reaches it) and data residency (whether sensitive data, and any training derived from it, stay within New Zealand’s legal reach). A pre-vetted list of suppliers that meet the tests keeps the buying process simple for agencies. The lever already exists. The Government Procurement Rules are issued by the Ministry of Business, Innovation and Employment (MBIE) and approved by Cabinet; the 5th edition took effect in December 2025 and is the natural vehicle for a new clause. The Government Chief Digital Officer’s Cloud-First policy already says agencies should, over time, store RESTRICTED data in a New Zealand-based data centre where a suitable onshore service exists, so the principle of a residency test is established in current policy and only needs extending to AI. Owner: MBIE (Government Procurement), with the Government Chief Digital Officer. Mechanism: a clause in the next refresh of the Rules, plus supplier-list guidance. Indicative cost: administrative only. Sequencing: first, because it is the rule everything else leans on.
A data-handling standard. This is a clear, plain statement of what classes of information may and may not be put into externally controlled AI tools, so that staff finally know where the line sits and the everyday leakage of sensitive information stops being invisible. Owner: the Government Chief Data Steward (within Stats NZ), with the Government Chief Digital Officer. Mechanism: a data-handling standard issued under the existing all-of-government data-system role, aligned to the procurement test so the two reinforce each other. Indicative cost: administrative only. Sequencing: alongside the procurement test.
Oversight you can check. Today’s guidance is voluntary: the 2020 Algorithm Charter and the 2025 Public Service AI Framework set expectations but bind no one and maintain no register. This phase makes oversight real for public-sector AI through three connected parts. First, a provenance-and-audit standard: a tamper-evident record of what a government system decided and on what basis. Second, a public register of significant government AI, so the use of these systems is a matter of record rather than discovery. Third, an accountable monitor with the standing to maintain the register and to audit. This monitor is the one genuinely new function this proposal asks for; everything else repurposes machinery New Zealand already has. It can be established in either of two minimal forms: a standalone monitor reporting to Parliament, or an expanded mandate for an existing officer. Both build directly on the Charter and Framework already in place. Owner: a new oversight-and-register function, anchored to the existing Stats NZ and Government Chief Digital Officer foundations. Mechanism: the provenance standard and register can be issued administratively; the monitor’s audit powers are best set in statute for durability, but an interim administrative mandate can operate from day one. Indicative cost: low; a small standing function rather than a new agency. Sequencing: stand up the standard and register first, confirm the monitor’s mandate in the Phase three review.
Indigenous data sovereignty, operationalised through existing instruments. Rather than author new consultation, this phase adopts the Māori-led data-governance instruments already in use for public-sector AI, sets a consent standard for AI trained on community data, and keeps indigenous data within New Zealand’s jurisdiction. This sits within existing Stats NZ data-sovereignty arrangements; the relevant Māori-led instruments are listed in the sources rather than restated here, because that ground is governed by Māori-led frameworks and this proposal does not speak for it. Owner: existing Stats NZ arrangements, working with the Māori-led bodies that hold these instruments. Mechanism: adoption of current instruments into the procurement test and data-handling standard. Indicative cost: administrative only. Sequencing: with the rest of Phase one.
The organisations most exposed to the risks in this proposal are the least equipped to manage them: small businesses, not-for-profits, trusts, councils, primary-sector operators, and many public agencies. They will not set up an AI department, and what the market offers them is largely foreign, often monetises their data, and is rarely built for organisations of their size. Phase two equips them. It is the one part of the pathway that needs real money.
A national AI capability fund. This is a fund with three uses: subsidised access to compute so these organisations can run AI under their own control; advisory support so they can adopt it well; and support for community- and sector-scale models, which give people a safe place for sensitive work and a constructive answer to the unsanctioned tools staff already use. Comparable small economies have done exactly this. Canada operates a public compute-access fund, Singapore pairs enterprise compute credits with advisory support, and Ireland runs national programmes for small and medium firms. The lever is a new Budget appropriation. Owner: Treasury provides the appropriation; MBIE administers it; delivery is routed through the new Research Funding New Zealand (RFNZ), into which the Marsden, Endeavour and Strategic Science Investment funds are being consolidated; Callaghan Innovation, a separate agency, is being disestablished around 30 June 2026, its functions redistributed. Mechanism: a contestable fund with compute, advisory, and model-support streams. Indicative cost: the substantive item in this proposal. Overseas programmes of this kind run from the low tens to the low hundreds of millions; scaled to New Zealand, an initial commitment in the low tens of millions over three to four years would be a credible start. These figures are indicative and benchmarked to the Canadian and Singaporean programmes; the precise figure is for Treasury and MBIE to cost. Sequencing: set up once the Phase one rules create demand for compliant capability.
Redeployment funding, so people move up rather than out. The fund should pay for people as well as tools. Where AI takes the routine part of a job, redeployment funding helps the people in that role move into better work rather than out of it. This is what turns adoption from a cost-cutting exercise into a productivity gain the public can see. Owner: MBIE, within the capability fund, drawing on existing skills and workforce programmes. Mechanism: a redeployment and reskilling stream tied to funded adoption projects. Sequencing: from the fund’s first year.
Reversible pilots. Public money in this phase backs pilots that can be corrected, not irreversible bets. Each funded project should be scoped so it can be stopped or unwound if the evidence turns against it, with that condition built into the funding terms. Owner: MBIE through the fund’s design. Mechanism: reversibility and a stop condition as a funding criterion. Sequencing: a standing rule across Phase two.
This phase puts in place the arrangements that keep the direction cross-party and let the system grow without becoming brittle.
The cross-party convening home. The direction needs a vehicle with standing across the House. A parliamentary select committee inquiry is that vehicle, and the natural host is the Economic Development, Science and Innovation Committee. The concrete first step is for a member to seek the inquiry. The committee’s technical work is supported by a Royal Society Te Apārangi expert panel, for which the precedent is its gene-editing panel, and by the Standards New Zealand SC 42 mirror committee, with the AI Forum’s convening reach available alongside. Delivery is then coordinated across agencies by an interdepartmental board, in the way the Digital Strategy for Aotearoa is already coordinated through its Digital Executive Board. Owner: the select committee for direction; an interdepartmental board for delivery. Indicative cost: within existing parliamentary and agency budgets. Sequencing: the inquiry can be sought now and run in parallel with Phase one.
Connect systems on open standards, not a central registry. Rather than route everything through one national database, separate systems should talk to each other directly using shared open standards: W3C decentralised identifiers and verifiable credentials, and the ISO/IEC SC 42 family. Connecting the parts this way keeps a failure in one place from bringing down the rest, and lets each authority, whether the Crown, an iwi, a sector regulator, a professional body, or the holder of the data, verify only what it is responsible for. There is no single point for an attacker to break through. Owner: the Government Chief Digital Officer for public-sector interoperability, with Standards New Zealand for the standards themselves. Indicative cost: low; standards adoption rather than a build. Sequencing: as systems are procured and renewed under the Phase one rules.
International engagement. New Zealand already mirrors the SC 42 standards through Standards New Zealand on a one-country-one-vote basis. This phase uses that seat actively, contributing to the standards that matter most here (42001 on AI management systems, 23894 on AI risk management, 23053 on the machine-learning lifecycle, and 22989 on terminology) and to the international indigenous-data-sovereignty networks New Zealand already engages with. Owner: Standards New Zealand. Indicative cost: modest membership and travel, within existing budgets. Sequencing: ongoing.
Built-in review. The direction should be correctable as the technology and the evidence change. A scheduled review, reporting to the select committee, keeps the settings current and confirms the standing arrangements, including the oversight monitor’s statutory mandate. Owner: the select committee, advised by the oversight monitor. Mechanism: a fixed review point written into the arrangements. Sequencing: first review at the end of Phase two.
The purpose of a pathway is to replace silos with a coordinated picture. The table below maps each lever to a realistic owner. With one exception, every owner is machinery New Zealand already has.
| Lever | Realistic owner | Existing instrument |
|---|---|---|
| AI strategy, procurement test, capability fund | MBIE (with Treasury for the appropriation) | Government Procurement Rules; Budget appropriation |
| Public-service AI, cloud and digital standards; interoperability | Government Chief Digital Officer (within DIA) | Public Service AI Framework; Cloud-First policy |
| Data system and data-handling standard | Government Chief Data Steward (Stats NZ) | All-of-government data-system mandate |
| Indigenous data governance | Existing Māori-led instruments, with Stats NZ | Māori-led data-governance frameworks (see sources) |
| Standards and international alignment | Standards New Zealand | ISO/IEC JTC 1/SC 42 mirror committee |
| Privacy | Office of the Privacy Commissioner | Privacy Act |
| Capability funding delivery | Research Funding New Zealand | Consolidated science funds |
| Cross-party direction | Parliamentary select committee inquiry | Economic Development, Science and Innovation Committee |
| Independent expertise | Royal Society Te Apārangi; academia | Expert-panel precedent (gene editing) |
| Delivery coordination | Interdepartmental board | Digital Executive Board model |
| Oversight, audit and public register | New oversight-and-register function | Builds on the Algorithm Charter and Public Service AI Framework |
The one genuinely new institution is the oversight-and-register function in the final row. It does not yet exist: there is no public register of government AI, the Algorithm Charter is voluntary, and the Public Service AI Framework is non-binding. This proposal asks for it to be created in minimal form, building on those two foundations rather than replacing them. The convening vehicle is the select committee inquiry, hosted by the Economic Development, Science and Innovation Committee and triggered by a single member seeking it. Everything else in the table repurposes an existing role.
The fiscal shape of this proposal is straightforward: almost all of Phase one and Phase three is rules-and-standards work that costs little, and there is a single capital item.
The low-cost work is the procurement test, the data-handling standard, the provenance-and-audit standard, the public register, the open-standards interoperability requirements, and the convening and review arrangements. None of these is a build. They are clauses, standards and committee processes delivered within existing mandates and budgets, which is why they can begin now.
The one substantive item is the national AI capability fund. On current benchmarks an initial commitment in the low tens of millions over three to four years would be a credible start. The costing basis is comparison with the Canadian, Singaporean and Irish programmes already running, which span from the low tens to the low hundreds of millions; the New Zealand figure should be set at the lower end of that range and costed precisely by Treasury and MBIE before any commitment. These figures are indicative, not a request for a specific sum.
Phasing keeps the fiscal exposure contained. The rules go first at near-zero cost. The fund follows only once those rules have created real demand for compliant capability, so money is spent against demand that exists rather than demand that is assumed. Delivery runs through Research Funding New Zealand, using consolidated machinery rather than standing up a new delivery body. Within the fund, every project is a reversible pilot with a stop condition, which limits the cost of any single bet.
For those who weigh every dollar, the value-for-money case is that the largest exposures here, the loss of control over critical systems and the daily outflow of sensitive information, are being created at no charge today and would be expensive to reverse later. The rules-first design buys most of the protection for administrative cost. The single fund is modest, benchmarked, phased, reversible, and delivered through existing channels. This is a restrained commitment, sequenced so that spending is pulled by demonstrated need.
Three principles govern how this pathway is run, and each is built into the arrangements rather than left to good intentions.
Human authority over consequential decisions. Decisions that materially affect people stay with accountable humans. Every automated decision path should have three outcomes, not two: proceed, refuse, or escalate to a human. What cannot be reduced to a rule must not be automated. This is a design requirement on government AI, checked through the procurement test and the audit standard rather than asserted as a principle.
Oversight you can check. The provenance-and-audit standard produces a tamper-evident record of what a government system decided and on what basis. The public register makes the use of significant government AI a matter of record. The accountable monitor has the standing to maintain the register and to audit against the standard, and reports to Parliament. Together these turn today’s voluntary guidance into oversight that an MP, an auditor or a member of the public can actually verify.
Reversibility, by design and by structure. Reversibility runs at two levels. At the project level, public money backs pilots that can be stopped or unwound, with the stop condition written into the funding. At the system level, connecting authorities on open standards rather than through one central registry means a failure or a capture in one place does not bring down the rest, and each authority verifies only what it is responsible for. The scheduled review then keeps the whole direction correctable as the technology and the evidence change. The aim throughout is custody that holds even when goodwill fails: protection that rests on how systems are engineered, procured and recorded, not on a distant assurance that can be withdrawn.
Māori data is a taonga. The obligations te Tiriti places on the Crown in this area are real, and they are best met not by writing fresh consultation into this document but by adopting the data-governance instruments Māori have already developed and recognised. Those instruments, led by Māori for Māori data, govern this ground; the relevant ones are named in the sources rather than restated here.
This document does not speak for Māori, and does not seek to. Its role is narrower: to make sure the wider direction it proposes leaves room for those frameworks to operate, and defers to them where they apply. In practical terms that means one firm commitment carried through every part of this proposal: indigenous data is kept within New Zealand’s jurisdiction and legal reach, so that authority over it remains where it belongs and cannot be exercised offshore.
Holding authority at home does not mean turning inward. The opposite is true: the surest way for a small country to keep control of its own AI is to build on the standards and partnerships its trading peers already use.
New Zealand already mirrors the international AI standards developed through ISO/IEC JTC 1/SC 42. These cover the shared vocabulary of AI, the machine-learning lifecycle, risk management, and AI management systems. Continuing and deepening that participation, through Standards New Zealand, costs little and keeps our rules legible to the rest of the world. It is the cheapest sovereignty money can buy: a seat at the table where the global ground rules are written.
Interoperability with the European Union’s AI Act matters for the same reason. The EU is a significant market, and its framework is becoming a reference point well beyond Europe. A New Zealand approach built to sit alongside it, rather than cut across it, protects market access for our exporters and spares them the cost of meeting two incompatible regimes. This is alignment as trade policy, not as deference.
New Zealand should also keep engaging with the networks of comparable small economies working through the same questions, and with the international indigenous-data-sovereignty networks in which it already takes part. Several small, open, trade-dependent democracies have moved ahead of us here, and there is more to gain from learning alongside them than from starting cold. None of this asks New Zealand to surrender any decision that is properly its own. It is how a country our size stays connected, credible, and in step with the markets and partners it depends on.
The direction set out here will draw familiar objections from across the House. Most rest on a misreading of what is, and is not, being proposed. We answer the main ones plainly.
Q. Is this protectionism, or anti-trade? A. No. It sets no tariff, blocks no supplier, and favours no flag. It asks two questions of the AI our public services buy: under whose law does it run, and where does our sensitive data go. Those are questions of jurisdiction and authority, not of trade barriers. The approach is built to be interoperable with the EU AI Act and with the international standards New Zealand already mirrors, which makes it easier, not harder, to trade.
Q. Can a small country afford this? A. Most of it costs very little, because most of it is regulatory work: a procurement test, a data-handling standard, and an oversight framework are rules and standards, not capital. The one item that needs real money is a capability fund to help smaller organisations adopt AI they control. Comparable economies run such programmes from the low tens to the low hundreds of millions; scaled to New Zealand, an initial commitment in the low tens of millions over three to four years would be a credible start, to be costed precisely against those benchmarks.
Q. Won’t this slow AI adoption? A. No. It is what makes safe adoption possible, particularly for the small businesses, not-for-profits, councils and primary-sector operators the market currently leaves behind. They are the most exposed and the least equipped, and what is on offer to them today is foreign, often monetises their data, and is rarely built for organisations their size. Giving them AI they can trust speeds adoption where it has been stuck.
Q. Why not just adopt AI, as the current strategy says? A. Adoption matters, and the 2025 national strategy is a sensible first step. But adoption on its own leaves the central exposures unaddressed: who controls the AI we depend on, where it runs, under whose law, and where our information goes when we use it. This is not a rebuke of that strategy. It is the necessary second step that completes it.
Q. Isn’t government picking winners? A. No. It sets standards and widens access; it does not choose suppliers or technologies. The procurement tests are open to any vendor that meets them, the capability fund is open to the organisations that need it, and the whole approach is deliberately vendor-neutral. Setting the rules of the road is the opposite of picking who gets to drive.
Q. Is this realistic? A. Yes. Comparable small economies already do most of it: public compute-access funds, enterprise compute initiatives paired with advisory support, national SME programmes. And almost everything proposed here repurposes machinery New Zealand already has, in procurement, data stewardship, standards and the select-committee system. Only one genuinely new function is asked for. This is assembly, not invention.
The ask is simple, and it is addressed to every party.
Convene a cross-party, multi-stakeholder process to set this direction. AI is one of the few questions on which New Zealand can act as one country, and it should not be left to split along the usual lines while the exposure grows.
The concrete first step is within reach of any member of Parliament: seek a select committee inquiry, with the Economic Development, Science and Innovation Committee the natural host. A committee inquiry gives the direction real standing, draws in expert and community evidence, and sets a cross-party course without any party surrendering what makes it distinct.
Alongside that, we invite the parties to stand together on the seven common-ground commitments set out earlier in this document. They are not left or right. They are the terms on which a small democratic country keeps hold of its own future, and few New Zealanders would reject any of them.
The substance is here, and the fuller framework that sits behind it is published openly. Both are offered freely, as a contribution to a decision that belongs to all of us, not as a product and not as a party document.
A closing word on scale and timing. The change AI is bringing is on the order of climate change in its reach into work, services and the decisions of everyday life. But unlike climate, we are still early. The infrastructure is only now being built, the rules are not yet written, and the choices are still open. That window will not stay open for long.
AI is not coming; it is already here. The only question left is whose terms it runs on — ours, or someone else’s. New Zealand can still make it ours: held under our own authority and our own law, and answerable to the people it serves. That choice is open now, and it will not stay open. If we act while we can, a small country at the edge of the world can remain the author of its own future. That is worth doing together, and it is worth doing now.
New Zealand has not yet set a clear direction on AI. If this makes sense to you, do two things: share it, and forward it to your MP — or to anyone who can help move it. Change here will come from enough people asking for it.
Share on X · Facebook · LinkedIn · Email it · Forward to your MP · Find your MP’s email
Comparator figures are drawn from official programme sources. The characterisation of the 12 June 2026 directive follows contemporaneous reporting; its precise legal designation is contested.
Author and declaration of interest: this draft proposal is offered by John G. Stroh (My Digital Sovereignty Ltd) as an independent contribution to public debate. It points to a working New Zealand platform that runs community AI on New Zealand and European infrastructure with no United States inference path — cited only as evidence that the approach is buildable. The proposal is written to stand on its merits and to be vendor-neutral; the author has a commercial interest in sovereign-AI infrastructure and declares it. This is a contribution to a public decision, not a procurement.