The pitch is simple. Replace people with machines. Save billions. Deliver faster services.

But when governments talk about using AI, the simplicity quickly evaporates.

In New Zealand, a plan to cut public service costs by $2.4 billion over four years by replacing staff with AI has exposed a deeper issue—not technological limits, but strategic confusion. When ministers were asked how exactly AI should be used, the answers were vague. So vague, in fact, that journalists turned to AI itself for clarity.

The irony is hard to ignore.

What emerged from those responses is not a vision of radical transformation, but a pattern of caution. Across different AI systems, the message is consistent: AI works best not as a replacement for people, but as a remover of friction.

And that distinction matters.

Because the temptation to treat AI as a blunt cost-cutting tool is strong—and dangerous. Governments operate in high-stakes environments where decisions affect people’s rights, finances, health, and futures. Automating the wrong layer doesn’t just reduce costs. It risks reducing accountability.

So where does AI actually move the needle?

Start with the most obvious: bureaucracy.

Across welfare agencies, tax authorities, immigration services and regulatory bodies, millions of processes are repetitive, document-heavy and predictable. Forms are checked. Data is entered. Files are summarized. Decisions follow rules. This is where AI thrives.

Automating these workflows doesn’t eliminate the need for public servants—it changes how they spend their time. Instead of processing routine cases, humans can focus on complexity: edge cases, exceptions, and human judgment calls.

In other words: less friction, more thinking.

The second major opportunity lies in policy work—the hidden engine of government.

Behind every public decision is a mountain of analysis: submissions, legal texts, historical precedents, impact assessments. Much of it is synthesis, comparison and drafting—tasks AI can already perform at speed.

Used properly, AI could compress weeks of policy preparation into days. Not by replacing expertise, but by accelerating everything around it. The cognitive load shifts from gathering information to interpreting it.

But this is also where misuse becomes tempting. If AI drafts the advice, who owns the judgment? Efficiency can quietly erode responsibility if governance doesn’t keep pace.

And then there is the frontline.

Here, the consensus sharpens: AI should assist, not decide.

In hospitals, schools, policing and social services, AI can summarize records, suggest next steps, translate information and reduce administrative burden. It can even help citizens navigate complex systems—finding the right benefit, the correct form, the appropriate agency.

But these systems should not replace human interaction in moments of vulnerability. A chatbot cannot assess trauma. An algorithm cannot weigh moral nuance.

This is not a technical limitation—it’s a societal boundary.

The real risk in the government’s current approach is not that AI will fail. It’s that it will succeed in the wrong way.

If deployed purely as a cost-cutting tool, AI could process more cases faster—but with less empathy, less discretion and less trust. Services might appear efficient on paper while quietly degrading in experience.

The alternative is harder, but smarter: design AI around value, not just savings.

That means shared platforms across agencies rather than fragmented experiments. It means strict guardrails around high-risk decisions. It means transparency about how systems are used—and where humans remain in control.

Above all, it means reframing the goal.

AI should not make the state smaller at any cost. It should make it sharper, faster and more responsive—without losing its human core.

Because in the end, the test is not whether AI can do the job.

It’s whether the public feels better served—or quietly pushed aside by the machines meant to help them.

‍