Peter Griffin: Agents of illusion? AI agents promise a smarter state but NZ is nowhere near ready to run one

At its May 2026 Google I/O conference in California, the company unveiled Gemini Spark, a so-called AI agent designed not just to respond to prompts but to act. It can plan, execute multi-step tasks, juggle different applications and operate with a level of autonomy that moves AI from assistant to operator. In Google’s framing, it’s less like search and more like staff, working away for you 24/7 even when you turn off your computer.

This is a significant shift that Google rivals OpenAI and Anthropic are also pursuing. We’re watching AI evolve from something that answers questions to something that gets things done. In demos, the systems booked travel, summarised documents, generated code and stitched together workflows across apps without constant human nudging. The promise is fewer clicks, fewer people, more output.

And it’s this promise that has caught the attention of New Zealand’s political class. The government’s plan to cut 8700 public sector roles has been justified, in part, by the idea that AI can pick up the slack, automating routine work, streamlining services and reducing the need for human bureaucrats.

It sounds neat. It also badly misreads where this technology is, and where New Zealand is starting from. Deploying AI agents at scale inside the government isn’t like installing Microsoft Office. It requires clean, well-structured data, modern interoperable systems, robust governance frameworks and a workforce that understands how to design, supervise and audit AI-driven processes.

New Zealand’s public sector has none of those things in abundance. What it does have is decades-old legacy infrastructure, fragmented databases, siloed agencies, and procurement processes that struggle to deliver even basic digital transformation on time and on budget.

Imagine dropping autonomous AI agents into that environment and expecting them to function reliably. The result wouldn’t be a sleek, automated state. It would be a mess, one where errors propagate faster, accountability becomes murkier and already strained services risk further degradation.

AI agents are only as good as the systems they sit on. In New Zealand’s case, that foundation is shaky. There’s also a deeper misunderstanding at play about what these tools do.

Yes, they can automate certain tasks: drafting documents, summarising information and handling predictable workflows. But much of the work done by public servants involves judgment, context, negotiation and accountability. It involves navigating ambiguity, not just executing instructions. Those are precisely the areas where AI still struggles.

We’ve seen this movie before. Grand technological promises used to justify sweeping structural change, only for reality to intrude later. Elon Musk’s much-hyped Department of Government Efficiency (DOGE) in the US, built on similar assumptions about automation and radical streamlining, has run into the hard limits of institutional complexity. Governments are not startups. They cannot simply “move fast and break things” without breaking trust, services and, ultimately, themselves. New Zealand risks importing that same mindset without learning the lessons.