In a meeting room on Lambton Quay, a paper is on its eleventh draft. It is about whether the public service should adopt AI. It concludes no.

On the wall, a screen is open to the agency’s approved AI model. The model performs like the free version of its frontier competitors. Nobody has asked it anything for some time.

There are eight people around the table: a privacy specialist, a subject matter expert, a manager, two directors, two deputy chief executives and someone whose role is hard to summarise but who has views. Each has left tracked changes and competing comments. The junior analyst, Hannah, will be working through yet another night reconciling them.

The group reasons that the juice is not worth the squeeze: studies on AI productivity show only modest gains, of 10 to 20 percent, drawn from settings unlike a policy shop. Yet the Public Service Commissioner, asked recently about AI adoption, said the Public Service has “lost its way”. Their paper does not mention this.
 
Hannah meets two friends for coffee. Sarah owns a salon and muses over having hired a second stylist on the back of the time AI has freed up. Mike is a financial adviser.
 
Sarah mentions Hannah’s paper and asks them about the government’s plan to cut 8,700 public servants by 2029, around 14 per cent of the workforce.
 
“10 to 20 percent productivity gain would deliver those cuts, no?” Mike says. “And most of those studies are about customer-facing stuff. I’m using it a lot more than that for analysis, even writing. AI drafts and redrafts my advice now, in minutes. I just review.”
 
“My books are basically doing themselves,” Sarah says. “That’s how I could afford Maria.”

Mike laughs.

“I was up last night redrafting my paper,” Hannah groans.

“I can’t believe you policy peeps are still manually writing papers,” Mike replies.

Hannah does not say anything.

Hannah is back in the meeting room. The paper is on its twelfth draft. The screen on the wall is still open. She has not been told whether she is allowed to ask it anything.

Late that night, working through the new tracked changes, she thinks about Sarah’s extra free time and Mike drafting and redrafting in minutes.

“AI could be doing this,” she thinks. After a moment, “Maybe I should go work for Mike.”

She opens a tab and pays for an AI account. The night’s reconciliation takes twenty minutes. The paper still needs a thirteenth review.