NZ Government AI and Decision Governance: When Copilot Isn't Enough
- David Harvey
- 5 days ago
- 12 min read
A response to Phil Nevell, Andrew Hamilton and to New Zealand's AI moment.
Phil Nevell posted something over a month ago that I haven't been able to stop thinking about.
Responding to Sam Altman's vision for the Intelligence Age, Phil asked the question that most AI commentary carefully avoids: who actually benefits here, and who just becomes more dependent?
He wrote: "People-first AI, in my view, means building genuine capability in individuals so they show up to this new world with skills and confidence, not as grateful passengers on someone else's infrastructure."
Andrew Hamilton responded with the sharpest analytical point in the thread. He identified the priority-ordering problem in OpenAI's own policy paper: prosperity first, risk mitigation second, democratised access third. His counter-argument was precise: "If instead the first move is getting people genuinely AI capable in an integrated economy, the displacement problem shrinks before it starts. The safety net stays smaller because fewer people need it."
Then Andrew landed the line that connects all of this: "I'd rather see policy that builds human-AI partnerships than one that builds a bigger safety net."
Phil's response was equally clear: "Critical for NZ Inc., we move the capability needle and avoid dependency. The role of strong Boards and company leaders has never been more important."
Between them, Phil and Andrew described exactly the system I have been building for four years. They named the problem, the policy failure, the right direction, and the right actors. What neither of them yet knows is that a production-level answer to this gap is already running in New Zealand.
Then, the same week, the New Zealand Government handed us the most concrete test case imaginable.
9,000 people. A fragmented AI strategy. No receipt.
The news landed cleanly: New Zealand plans to reduce the public service by approximately 14% — around 8700-9,000 roles — to return the civil service to its historical baseline of roughly 55,000. Finance Minister Nicola Willis and the coalition government are championing AI to automate administrative and processing tasks to cover these capacity reductions.
Nicola Willis has mentioned her department's use of Microsoft 365 Copilot, though that was no endorsement of a single government-wide AI governance framework. The actual picture is considerably more fragmented, and that fragmentation is the argument.
MBIE — the Ministry of Business, Innovation and Employment — previously banned AI tools altogether due to data and security concerns. They have since deployed Microsoft 365 Copilot for internal productivity work. ACC uses it to summarise documents and reduce the time processing ministerial reports. Parliamentary Service allows MPs and Beehive staff to use ChatGPT, with restrictions on the use of confidential information and a mandate to fact-check outputs.
Health New Zealand has gone the other direction entirely: staff are strictly prohibited from using free AI tools, including ChatGPT, for clinical notes or medical records. The stated reasons are data privacy and accountability concerns.
The Department of Corrections prohibits staff from entering private offender information into AI tools or using them for sensitive assessments.
Read that back slowly. The same administration that bans AI from clinical notes due to accountability concerns is simultaneously using AI productivity tools as part of the capacity-replacement architecture that will end 9,000 public servants' careers.
The issue is not that Copilot personally chooses which roles disappear. The issue is that AI-enabled productivity is now being used to justify why those roles are no longer needed. That makes decision governance non-optional. If AI helps justify, absorb, sequence, or administer a workforce reduction, then the public deserves more than a productivity claim. It deserves a decision receipt.
The contradiction is not a failure of intent. It is a failure of infrastructure. There is no governed decision layer sitting between the AI outputs and the consequential choices. There is no receipt. There is no audit trail that survives external scrutiny.
I want to be precise here, because precision matters when 9,000 people's livelihoods are at stake.
Copilot and ChatGPT are probabilistic systems.
That is not a criticism. That is a description of how it works. Large language models generate statistically likely outputs. Given a prompt, they produce the most plausible continuation based on patterns in training data. That is genuinely useful for drafting documents, summarising meetings, and accelerating routine knowledge work.
Use Copilot where productivity is the task. Use governed receipts where accountability is the requirement.
NZ Government AI & Decision Governance
The problem is not that the government is using AI. The problem is that it has no governing layer above the AI to ensure that the decisions the AI informs can be audited, evidenced, and held to account.
Decisions about retrenching 9,000 people over two years are not productivity tasks. They are deterministic governance requirements.
They require an evidence trail. An audit record. A chain of reasoning that survives an Office of the Auditor-General inquiry, a Select Committee hearing, or a judicial review. They require someone to be accountable — not for what the model generated, but for the decision that was made.
A probability distribution doesn't do that. A receipt does.
The question nobody in Wellington is asking out loud
When the OAG comes asking — and eventually they will — the question will be straightforward: "Show us the evidence trail for how this decision was made."
Copilot produces an output. It does not produce a governed, LAW-compliant, longitudinally auditable decision record. It does not generate a Comfort Index before and after the deliberation. It does not create a receipt that survives external scrutiny with the decision-maker's reasoning intact and time-stamped.
This is not a gap Microsoft is motivated to close. Copilot's value proposition is productivity — doing more, faster, at lower cost. Governance accountability is a different product category entirely. And it is the product category that New Zealand actually needs right now.
What we built — and why it's running in production, not in a pitch deck
Last year, in conversation, Phil told me, "You need a Government or Enterprise client."
He was right. And I've been building the answer ever since. We have a pilot Clinician client, but something more structurally interesting: a governed cohort study.
Ask Omega* (by Design By Zen) is a governed decision-intelligence framework built on the same clinical-grade evidence architecture as SHE ZenAI — designed specifically for high-stakes decisions where evidence, not likelihood, is the currency of accountability.
Every decision processed generates a LAW-compliant audit trail, a Comfort Index delta, a recoverable decision thread, and a receipt that belongs to the person who made the decision — not to the platform that helped them make it.
The difference from Copilot is architectural, not cosmetic.
Copilot is built on a probabilistic foundation optimised for output generation. Ask Omega* is built on a deterministic governance calculus — Q*(E8) — designed specifically for high-stakes decisions where evidence, not likelihood, is the currency of accountability.
And it is running. Not in a demo environment. In production, with X402 billing, live LTMS memory traces, real checkpoint timestamps, and a SKU catalogue that charges $7 per governed decision. The receipts are real. The audit trail exists. The Comfort Index moves over time. It's human-aligned by design.
"I'm not aware of a current local AI provider." Minister, the category now exists.
Labour asked in Parliament what the rollout and licensing cost of the government's AI strategy would be? Digitising Government Minister Paul Goldsmith replied that he didn't have the exact figure, that it varies, and that the government is working toward "a more coherent and centrally guided system."
When asked whether it would use local or overseas AI technology, Goldsmith said: "I'm not aware of a current local AI provider in the scale of Claude or Copilot, but what I would say is that we'll be making use of the best technology available."
That statement deserves a precise response.
New Zealand does not need a local large language model at the scale of Claude or Copilot. Minister Goldsmith is right about that. That is not the missing category. The missing category is a New Zealand-governed decision layer that can sit above any AI tool — Claude, Copilot, or anything that follows — preserve evidence, record human accountability, and survive domestic scrutiny.
New Zealand does not need to own every model it uses. But it does need to own the decision governance layer above those models.
That layer already exists. It was built in New Zealand. It runs in production. It is not for sale to Anthropic or Microsoft.
Minister, the category now exists.
The cost the published material doesn't show
Professor Alexandra Andhov, chair of law and technology at the University of Auckland, told RNZ something that deserves to sit alongside the $2.4 billion in savings cited as justification for 9,000 role eliminations.
"The published material doesn't really show the cost side of the AI. - Professor Alexandra Andhov"
Enterprise-scale AI is not a one-off purchase. It entails ongoing licence fees, model upgrade costs, and the compounding dependence on letting a US commercial entity decide when and how its tools are replaced. Then there is the pricing trajectory that every academic who studies this market understands, and that no government procurement document is currently accounting for.
"The costs that we pay for AI today are heavily subsidised while the AI companies are trying to capture as much of the market," Andhov said. "These are not the real costs that AI will cost."
When the subsidy ends — when OpenAI and Microsoft have captured sufficient market share and need to return their infrastructure investment — the pricing will shift. The government will have built its public service around tools whose cost structures it has not modelled and whose pricing it does not control.
That is what Silicon Valley quicksand looks like from the outside.
The hidden cost is not only licence fees. It is an architectural dependency. When workforce redesign, productivity claims, and public-sector decision accountability are routed through tools the Crown does not own, cannot fully audit, and cannot price-control over time, the savings become harder to prove. The question is not whether Copilot can summarise a document. It can. The question is whether the decision architecture for the 8,700 lost roles can produce a governed record that withstands scrutiny after the savings have been booked.
Andhov then named the loss that compounds all the others: "If the New Zealand government ultimately uses Microsoft's AI... then they're paying OpenAI, which is based in California, which doesn't pay any taxes here. All of this amount is taken to the US and actually brings nothing back to New Zealand... and it has involved loss of jobs here."
Count the losses: circa 9,000 public sector roles eliminated. The savings that were meant to replace them are flowing to California. No NZ tax revenue from the AI that caused the displacement. No local capability built. No governed accountability layer for the decisions that created all of the above.
That is not a technology strategy. That is a sovereign dependency being written into legislation and called efficiency.
Built in New Zealand. Governed by New Zealand law. Not for sale.
Andhov asked the question the procurement process has not formally answered: "Who are these AI providers? The majority of the providers that the government is considering are not New Zealand companies, not companies that are governed by New Zealand law, but US-based companies that only need to comply with US law."
New Zealand does not need to own every model it uses. But it does need to own the decision governance layer above those models. That is the thesis. That layer must be subject to New Zealand law, governed by New Zealand institutional interests, and auditable by New Zealand oversight bodies — the OAG, Select Committees, and the courts.
Ask Omega* is built here. The IP is unencumbered. It is not subject to US commercial law, US data jurisdiction, or US pricing decisions. When Node-1 — the local appliance that delivers governed decision intelligence from mobile to on-premises hardware — deploys, the decision data will not leave the jurisdiction where it was made.
Minister Goldsmith told David Seymour that New Zealand should "focus on the things we do well and sell them to the world."
That is precisely the argument. Governed, evidence-based decision intelligence is something New Zealand can build, own, and export. It requires no local silicon fabrication.
It requires no local LLM at Claude's or ChatGPT scale. It requires building an accountability layer that sits above those tools and holds the humans who use them accountable to the people they serve.
That is not a patriotic argument. It is a structural one. Sovereignty over your decision governance infrastructure is not optional when the decisions being made affect 9,000 people's livelihoods.
The governed cohort study — the answer to Phil's challenge deserved
The DBZ 1000 is our proposed mechanism. Enterprise and Government bodies don't become traditional clients — they become study sponsors. They fund cohort slots. They receive aggregate decision intelligence reports. They get the governance narrative they need before the accountability question becomes a political crisis.
For the NZ Govt specifically: the decisions being made right now, about which roles go and which don't, are the decisions that need a governed evidence trail most urgently. An evidence trail created after the fact is worthless. The receipts need to exist at the moment of greatest consequence: the next 12 months.
One thousand participants. 2026 to 2030. Health, wealth, work, relationships, connectivity — and now, explicitly, governance. Each participant generates a recoverable decision trail, measurable Comfort Index movement, and four years of longitudinal audit continuity.
The tool that displaced them offers them nothing. Ask Omega* could.
Here is the dimension of this story that nobody is discussing.
The 9,000 people being made redundant by this AI transformation are not, in the main, people who lack decision-making capability. Many of them have spent careers doing exactly that — structured analysis, evidence assessment, stakeholder navigation, policy reasoning. Their professional expertise is "governed thinking". That is precisely the skill that the AI era claims to need more of, not less.
Microsoft Copilot is simultaneously the cited mechanism for eliminating their roles and a tool that offers them nothing whatsoever on the other side of that elimination. It does not help a senior policy analyst at MBIE figure out what comes next. It does not help them navigate the most consequential decision of their professional life, "what to do now" with evidence, structure, and a governed record of their reasoning.
Ask Omega* does.
The 7-Day Comfort Index Challenge is not just a consumer onboarding flow. For someone facing a sudden and unwanted career transition, it is a structured way to surface what they actually value, where their discomfort lives, and what decisions are looping without resolution. That is clinical-grade support for a genuinely clinical-grade life event.
But the opportunity extends beyond support. These experienced decision-makers have irreplaceable domain expertise — deep institutional knowledge in health policy, economic analysis, social services, and infrastructure governance. That expertise, combined with Ask Omega*'s governance engine, could become a profession in its own right.
Decision Governance Practitioner.
A human expert with domain knowledge, using the Ask Omega* platform to help others create governed decision receipts in their field. Not replacing AI. Not being replaced by it. Working alongside it as the evidence layer that makes the output accountable.
Phil's question — who actually benefits here? — has a specific answer in this context. A senior MBIE analyst who loses their role to AI in 2026 could, with the right platform and framework, become a governed decision practitioner in the same domain by 2027. Their expertise is the IP. Ask Omega* is the infrastructure that makes that expertise commercially viable in the AI era.
Copilot cannot offer that. It is not designed to. Ask Omega* is.
The wrong priority ordering — and why capability must come first
Andrew's critique of the OpenAI paper's sequencing is worth dwelling on, because it applies directly to what is happening in Wellington right now.
Prosperity first, risk mitigation second, democratised access third. Andrew's counter is the more honest architecture: capability first means fewer people need the safety net because fewer people are helplessly displaced to begin with. The safety net stays smaller because the capability was built before the displacement arrived.
He also named the control problem that the Public Wealth Fund conversation is carefully avoiding: whoever controls the redistribution mechanism controls the conditions on which it is distributed. If the fund is administered by the same interests that control the AI creating the displacement, you have not distributed power. You have centralised it twice.
Ask Omega* is the capability-first model in practical form. Not a safety net. Not a redistribution mechanism. A governed decision platform that builds genuine capability in individuals — decision-making skill, evidence discipline, structured reasoning — so they arrive at the AI era with something to offer, not just something to lose.
Andrew's closing line is the architectural brief for what we built: "I'd rather see policy that builds human-AI partnerships than one that builds a bigger safety net."
Ask Omega* is a human-AI partnership at the level of each individual decision. The human brings the context, the stakes, and the values. The AI governs the process. The receipt belongs to the human. The reasoning is the human's. Accountability rests with the person who made the call — not with the platform that helped them make it.
Phil added the actor: "The role of strong Boards and company leaders has never been more important."
Agreed. And strong Boards making high-stakes decisions — about AI transformation, workforce restructuring, strategic pivots — need governed decision receipts rather than probability distributions. That is precisely what Ask Omega* produces. That is precisely what ExecDecide, our derivative tool for C Suite coaching, is being built to deliver.
To Phil and Andrew — publicly
Phil, you asked who is actually first in the Intelligence Age. Andrew answered it more precisely than most policy papers do: capability first, partnerships, not passengers, human accountability, not distributed dependency.
My contribution is the production-level answer to the question you both asked. The system is running. The receipts are real. The Comfort Index moves. The billing is live. And the cohort study that will generate four years of longitudinal evidence for governed human decision-making starts this year.
Phil — I have one question for you before I show you the rest: who do you know in Wellington who needs to see this before the accountability question becomes political?
Andrew — the human-AI partnership architecture you described is not theoretical here. It is in production code, built in New Zealand, not for sale to the people who built Copilot.
Call this week.
Omega* Sensing is part of the Omega* Unified Ecosystem, developed by Design By Zen, an NZ-based AI Lab. Omega* is the algorithmic engine beneath the ecosystem. SHE ZenAI is the brand of a governed clinical intelligence framework designed for high-trust domains where evidence, not confidence, is the currency of care. Version 1.0, April 2026.
Comments