Replacing jobs with AI. Does it work?

NZ Government plans to cut 9,000 public sector jobs by 2029, merge agencies, and expand AI to fill the gap. We are not alone. The UK, Canada, and the US have all announced the same bet: cut headcount, deploy AI, bank the savings.

So I went looking for a country where this has worked. Here is what I have found so far.

The US cut federal employment by 271,000 in 2025, a 9% decline. Spending did not fall. It rose. Around 25,000 people were rehired as essential, including nuclear security staff brought back within 48 hours. The IRS lost most of its senior tech leadership, including associate chief information officers and cybersecurity experts that were supporting agency's overarching IT modernisation efforts. Federal agencies were also forced to backfill vacant government roles with private contractors, which cost on average 1.83x more than standard federal employee rates, and nearly 3x more for Department of Defence roles. 

The UK tied AI to headcount reduction in 2024 and promised billions in savings. By December 2025 the civil service had grown 1.2%.

Estonia, the model everyone cites, never cut headcount through digitisation. Its public sector employment stayed flat and sits above the OECD average today. Estonia transformed service quality, not payroll.

The only documented wins are narrow and task-level. A Swedish municipality automated welfare reapplications and redeployed caseworkers to face-to-face work, which improved public satisfaction levels. The UK built a tool that analysed 50,000 consultation responses in two hours. In both cases, staff time was reallocated, not removed as their time was freed up to deliver more meaningful work. 

The pattern is consistent. Countries that invested in foundations first got real but modest gains in productivity. Countries that cut first and expected AI to backfill got rehiring, service degradation, and higher costs at the end of the day.

I have co-led technology simplification inside NZ government. The savings were real. They also took dedicated teams, sustained funding, and time before they appeared.

The goal is achievable. The sequencing decides whether we get Estonia's result or the American one.

Which examples have I missed? I want to know if this has worked somewhere.