October 1, 2026

Ministers chose a 55,000 headcount before pricing the AI meant to deliver it

43rd annual West Point Debate Tournament

Shrinking the core public service to 55,000 is a defensible goal. Staking that goal on artificial intelligence that nobody in Wellington has costed is a much weaker plan. Documents released under the Official Information Act show the target came from ministers first, with the technology meant to deliver it treated as a detail to sort out later.

For business owners who have ever signed off a software rollout, that sequence should ring alarm bells.

The number came before the plan

In April, three weeks before the Budget, Public Service Minister Paul Goldsmith told officials to “think about 55,000 as a starting point”. Officials had been modelling 60,000 by 2029. They warned the sharper cut could worsen performance risks and suggested pushing the deadline out to 2031.

Public Service Commissioner Sir Brian Roche sided with ministers within days, calling 55,000 a minimum and saying hitting it would need a much more aggressive approach to AI. He also conceded, with admirable candour, that “we have tried this before with very limited success”, and RNZ’s review of the papers found little detailed modelling of implementation, savings or the cost of using more AI.

Nicola Willis’s Budget package commits to cutting roughly 8700 roles and saving $2.4 billion by July 2029. Agencies are being asked to propose mergers built around citizen-facing functions on common technology platforms, with the detail still to come.

The case for cutting is solid

None of this means the bureaucracy is lean. Treasury’s Fiscal Strategy Report shows the public service grew from roughly 47,000 to more than 65,000 people between 2017 and 2023, with back-office functions growing faster than frontline delivery. New Zealand runs 39 departments and ministries administering Budget lines, against 16 in Australia and about 12 in Finland. That is a structure that begs for consolidation.

There is also evidence the government can execute on savings. In 2024, the public service shed an estimated 1723 roles through redundancy in the year to September, and that earlier round beat its operating savings target.

Headcount is still going the wrong way

The problem is that the machine keeps growing even as ministers talk it down. The latest workforce data shows 64,535 FTEs in the March 2026 quarter, up 1.4 percent, or 877 people, in three months. There is some good news in the same release: 23 of 40 departments shrank, and contractor spend fell to 5.3 percent of workforce costs, so agencies are not simply swapping staff for consultants.

Still, getting from 64,500 to 55,000 in under three years means removing close to one in seven roles, at a pace officials said has not been seen since the reforms of the early 1990s.

AI is a cost line, not a slogan

When Labour asked what the AI rollout and licensing would cost, Goldsmith told Parliament “I don’t have that exact figure at the moment”. Enterprise AI carries recurring licence fees, vendor-driven upgrades, audit and oversight, and cyber security spend that grows with every system connected to it.

University of Auckland Professor Alexandra Andhov, who chairs law and technology, puts it bluntly: “The argument that ‘we’ll use more AI’ is not a budget line. It is a major policy decision.” She argues current AI pricing is subsidised by vendors chasing market share and will rise.

Other academic analysis warns that restructuring agencies, shrinking staff and scaling AI all at once risks draining institutional knowledge exactly when oversight demands increase. That is a fair warning, not an argument for doing nothing.

What a real business would demand

Roche defended the approach by saying that if it were his business and his money, this is what he would do. Fair enough, but no board would approve a 14 percent workforce cut justified by technology without a business case showing what that technology costs, which processes it replaces and when the savings land.

For firms dealing with IRD, MBIE or MSD, the stakes are practical. If headcount falls before automated systems work, processing times, licensing, consents and compliance queries slow down, and the cost of that delay lands on the private sector. For technology suppliers, the flip side is a large procurement pipeline, provided agencies know what they are buying.

The test starts now

The government is right that Wellington is overstaffed and under-productive. But a smaller headcount is an input, not an outcome. The measure that matters is whether services get faster and cheaper per transaction. Ministers should publish the AI costings, agency by agency, before the next round of budget cuts bites. Without them, the $2.4 billion is a hope, and taxpayers have funded enough hopes already.

Sources

Reader Poll
Community

Join the discussion

Add useful context, ask a good question, or challenge an idea — keep it specific and respectful.

Create a commenter account

Enter the name you want shown publicly and your email. We will email you a password-set link; you cannot comment until you use it.

Your email is used for sign-in and account security. It is not published with comments.

Subscribe for weekly news

Subscribe For Weekly News

* indicates required