The New Zealand government has announced a reduction of roughly 9,000 public service roles, with the explicit intention of replacing much of that capacity with AI. The accompanying reporting from Stuff is worth reading in full, because it captures something we don't often see stated so plainly: ministers being asked what they think AI actually does, and most of them not really knowing.
We build AI systems for a living. We are not opposed to AI replacing work — a meaningful portion of what we ship at Dreamware does exactly that. So this isn't a piece arguing that the public service shouldn't change. It's a piece arguing that the order of operations matters, and that the current approach has the order exactly backwards.
The textbook way not to do this
The pattern goes like this: decide on a headcount reduction, announce it publicly, and then work out how AI will fill the gap. This is the wrong sequence for three reasons, and they compound.
First, you've fixed the wrong variable. The number of people doing a job is an output, not an input. It's downstream of what the work is, how it's structured, what tools exist, and what quality bar you're holding. When you fix headcount first and let AI capability be the unknown, you guarantee one of two outcomes: either the AI doesn't reach the required capability and service quality drops, or the target gets quietly missed and credibility erodes. There's no third path where it just works out.
Second, you've removed your own option to learn. AI adoption in any large organisation is iterative. You pilot, measure, adjust, expand. That cycle requires the people who currently do the work to be involved — because they know which parts are tedious and rule-based (good AI candidates), which parts require judgement and context (bad AI candidates), and which parts look like one but are actually the other (the dangerous middle). Cutting those people before the pilots run means you've thrown away the institutional knowledge that would have told you what to automate.
Third, you've created adversarial incentives. If your job is going to be replaced by AI, you are not going to enthusiastically help design the AI that replaces it. This is human, not malicious. The most successful AI rollouts we've seen — in private sector clients, but the principle generalises — have involved the people whose work is changing, with explicit commitments about what happens to them. "You will keep your role, but it will look different" is a workable conversation. "You are being made redundant, please train your replacement" is not.
What the ministers actually said
The Stuff article is instructive because it asked ministers, individually, what AI could do in their portfolios. The answers ranged from vague ("efficiencies") to slightly concerning (suggestions that AI could handle frontline social services), with very little technical specificity. One minister suggested AI could process consent applications. Another mentioned chatbots.
These aren't wrong, exactly. AI can absolutely help with consent processing — we've built systems that do similar work. But "AI can help with consent processing" and "AI can replace the people doing consent processing" are radically different claims. The first is a tooling decision that might make a team 30% more productive. The second is an architectural claim about end-to-end automation that requires solving document understanding, exception handling, audit trails, legal accountability, and integration with legacy systems that were never designed for machine consumers.
The gap between those two statements is where most government AI projects fail. Not because the technology can't do it, but because the procurement, the data, the change management, and the risk frameworks aren't ready.
What good adoption actually looks like
If we were advising on this — and to be clear, we're not — the sequence would look roughly like this.
Start with a capability audit, not a headcount target. Pick a small number of agencies and map their work into three buckets: highly structured and rule-based, mixed judgement and structure, and inherently human. Be honest about which is which. Most agencies will be surprised at how little falls cleanly into the first bucket.
Then pilot. Real pilots, with measurable baselines, run for long enough to see the failure modes. AI systems perform well on demos and badly on edge cases, and the edge cases are where government work lives — because government, almost by definition, handles the cases the market doesn't. A chatbot that answers 80% of queries correctly is impressive in a demo and catastrophic in a benefits context, where the 20% it gets wrong are the people who most need accurate information.
Then, and only then, talk about workforce. Some roles will genuinely shrink. Others will grow — you need more people doing oversight, model evaluation, exception handling, and the messy human work that AI surfaces rather than eliminates. Net headcount might drop, but the shape of the workforce changes more than the size does. Pretending otherwise is how you end up with a service that's cheaper on paper and worse in practice.
The infrastructure problem nobody mentions
There's a quieter issue underneath all of this. New Zealand's public sector runs on a patchwork of systems, many of them old, poorly documented, and held together by the institutional knowledge of the people being made redundant. AI doesn't run on vibes. It runs on data — clean, accessible, well-structured data with clear lineage and governance.
The agencies we'd consider AI-ready in New Zealand are a small subset of the whole. Most are still working through basic data infrastructure: consolidating systems, fixing master data, sorting out who owns what. None of that work goes away because you've announced an AI strategy. It gets harder, because the people who knew where the bodies were buried have left.
This is the part that worries us most. The announcement frames AI as a substitute for people. In reality, for the next five to ten years, AI in government is going to require more skilled people, not fewer — data engineers, ML engineers, evaluation specialists, governance leads, domain experts who can tell you when the model is wrong. These people are expensive and scarce. Cutting 9,000 generalist roles doesn't free up budget for them in any obvious way, because the savings are promised back to the taxpayer.
Our actual view
AI will change public service work. It should. There is genuine waste in any large bureaucracy, and there are categories of work — document processing, triage, summarisation, routine correspondence — where modern models are genuinely capable and the productivity gains are real.
But the way to capture those gains is to do the work: pick the use cases, build the pilots, measure the outcomes, redesign the roles, retrain the people, and let headcount shift as a consequence. The way to fail is to announce the headcount reduction first and treat AI as a budget line item that will sort itself out.
The announcement, as it stands, is the second thing. We hope the execution is the first.