Where Do Our Tax Dollars Go: $2.4 Billion Savings, No Strategy No Plan
Note: This piece draws on my professional experience in process transformation and on the public sources cited throughout.
On 2 July, Sir Brian Roche released a review he commissioned into digital delivery across the New Zealand government. Three reviewers, Adrian Littlewood, Justin Gray and Matt Crockett, spent three weeks talking to the people living inside the system, agency chief executives, technology leaders, vendors, and staff who had worked the delivery side of major transformations. What they heard may have surprised the people conducting the review. It would not have surprised anyone who has been saying it for a year.
One agency leader told the reviewers, "in reality they just helped us admire the problem. It didn't make our lives any easier in delivering our project." Another said, "I didn't feel that GDDA had the confidence or the capability to help broker an answer to building a common system for use across agencies, so we built our own." A third put it plainer still, "the process became the focus rather than the outcomes we were trying to achieve." This is a central digital function that costs $42 million a year to run, with 170 people inside it, and the agencies paying for it through their own budgets are saying this to an independent reviewer, on the record.
I have been telling this story from inside one programme, a Cabinet mandated effort that promised $3.9 billion in savings and spent twenty months producing nothing it could show for it. A government review has just told the same story again, from outside the whole system, finding a central digital agency that costs $42 million a year and still cannot tell the agencies paying for it how to deliver anything. Both come down to the same gap, between digitising a process and standardising it.
Every industry that has ever automated successfully closed that gap first, long before any machine was trusted to run it. Picture a small workshop building its first car by hand. Nobody knows yet, on day one, how long the door hinge should be, or which order the wiring goes in, or how much torque the wheel bolts need. They find out by doing it, badly at first, then less badly, adjusting the process every time something rattles loose. After enough cars, the steps stop changing. That is standardisation, and it only exists because someone did the work of building the same thing over and over until the variation disappeared.
Optimising comes after standardising, and it matters just as much. Once every car comes off the line built the same way, the workshop spends a further stretch of time refining it, removing a screw here because two were never needed, pre fitting a panel there because the old sequence wasted forty minutes a car for no reason anyone could defend. Optimisation is only possible once a process is stable enough to see the waste in it clearly.
AI is what governments now propose to put at the end of that sequence, and it behaves nothing like the optimisation step that is supposed to come before it. AI systems are trained on data, meaning they are shown enormous numbers of examples of how a process has actually been carried out in the past, and they build a model of what to do next from the patterns in those examples. This is often described as the system learning, and that is a fair word for what is happening mechanically. But it matters enormously what the system is learning from. It only has access to what the process already was, recorded in whatever data exists about it. If that data includes every workaround a stretched team invented under pressure, every exception made for convenience, every inconsistency that built up because nobody standardised the process in the first place, the system has no way to flag any of that as an error. It learns those inconsistencies as if they were the process, because as far as the data is concerned, they are.
The dangerous part is not the wrong answer itself. It is that the system delivers the wrong answer with the same fluency and the same apparent confidence as the right one, and confidence is very often mistaken for correctness. A person doing an inconsistent process badly can usually be asked where they were unsure. A trained model carries no such uncertainty forward. It cannot tell the difference between the two, so it does not hedge.
Two governments have already run this experiment at national scale, automating decisions over real people's lives before the process underneath them was ever standardised, and both produced the same failure. In Australia, Robodebt calculated welfare debts by averaging a person's annual income and comparing it against fortnightly payments, treating people as though they had received money in weeks when they had not. A Royal Commission later found the scheme had been pushed through despite legal advice warning it was unlawful. It wrongfully recovered AU$746 million from 381,000 people. The government wrote off $1.75 billion in debts once the scale of the error became undeniable, and has since paid more than $587 million in compensation across two class actions. The Commission heard from mothers whose sons died by suicide after being wrongly pursued for money they never owed, and said it believed theirs were not the only such losses connected to the scheme.
The United Kingdom ran its own version of the same experiment earlier and is still paying for it. Accounting software called Horizon told the Post Office that money was missing from branches run by ordinary sub postmasters who had balanced their own tills for years without incident. Nearly a thousand people were prosecuted on the strength of what the software reported. Some went to prison for theft a piece of software had invented. Compensation costs are now expected to exceed one billion pounds, and the country is still working through claims more than twenty years after Horizon was first installed. Both schemes were sold as savings. Neither government budgeted for what it would cost, in money or in people harmed, to automate a process that had never been verified first.
That is the weight sitting underneath New Zealand's current plan. Finance Minister Nicola Willis has said the public service overhaul, including greater use of AI, will deliver $2.4 billion in savings over four years, alongside close to 8,700 fewer roles by 2029. When Labour asked in Parliament what the rollout and licensing costs of that AI would actually be, Digitising Government Minister Paul Goldsmith did not have a figure to give them. Professor Alexandra Andhov at the University of Auckland has publicly questioned how the $2.4 billion number was reached at all, since the published material does not show the cost side of the technology.
Unlike the $2.4 billion figure, two agencies inside government already have a verifiable version of this done properly, sitting in the same review that catalogues everyone else's fragmentation. Kāinga Ora self funded a technology upgrade worth roughly $170 million by doing the unglamorous work first, low tech process and cost improvements made before any new system went in. Inland Revenue's transformation is a narrower kind of success than it is sometimes given credit for. Three of its ten targeted outcome measures were not achieved, and the Ombudsman has separately found its child support procedures deficient in how thoroughly they scrutinise a parent's finances before approving a payment arrangement. But the platform migration itself, replacing a 1980s mainframe and decommissioning more than four hundred legacy applications, delivered on time and $120 million under budget, in staged releases over several years, with stable funding and one accountable executive team throughout. Ninety nine percent of IR's transactions are digital today. Different cases, same underlying discipline. Both agencies did the unglamorous work, cost, process, or governance, as part of delivering the project, rather than skipping it in favour of moving faster. The rest of government has neither a number nor a working platform to point to, and the review says why directly, it was hard to get fundamental metrics to judge performance of tech investment across the agencies at all.
There is a version of this that works, and government does not need to invent it, because it already runs a working example of the same idea for a different problem. Before any system holding public data goes live, it has to pass a formal security check first, done by someone outside the team that actually built it. That rule exists because government learned, the hard way, that a team cannot be trusted to mark its own homework when the stakes are this high. Automation needs exactly the same kind of check, and currently has none. Before any agency is allowed to put AI onto a process, someone from outside that agency should have to look at real evidence, not a promise, that the process has actually been mapped the way that workshop mapped its car, step by step, that every agency using it does it the same way, and that someone has worked out, properly, what it will cost, rather than guessing at a number that sounds good in a press release. Until that evidence exists, and someone with nothing to gain from the answer has checked it, the automation should not go ahead, in exactly the same way a system that has not passed its security check is not allowed to hold the public's data. None of this asks government to build something new. It only asks government to apply the rule it already trusts for security to the thing it has never thought to apply it to.
The $2.4 billion this overhaul is meant to save has already been promised elsewhere, to health services, education, infrastructure, and the defence force and police. Robodebt and Horizon show what happens to a savings figure like that when the process underneath it was never verified. It does not disappear. It becomes a compensation bill, paid years later, larger than the number that was announced, drawn from the same public purse the hospitals and classrooms were waiting on. That is the actual decision in front of New Zealand right now, whether this money reaches the services it was promised to, or whether it ends up paying for a failure the country had every opportunity to see coming first.
