Recently the New Zealand Government announced their plans to cut the number of public servants & use AI to boost productivity in the public service. The response was negative - to put it politely.

But, what if AI can be used in the public service, & what would that look like? Let’s find out.

The Scenario

Let’s walk our way through a scenario. For this exercise I’m going to use the following issue: 

“Why are students dropping out of high school & how can we improve their life outcomes?”

Right out of the gate I need to state that some people drop out of high school & lead fulfilling lives packed with achievements. But on the flip side for some people, dropping out of high school is a substantial step towards an unfulfilling life, which has negative consequences on the individual & society.

Getting Started with AI

So, let’s use AI to address this important issue. But first, we’ve got two issues to address:

Issue 1: AI Data Sovereignty

Current consumer AI models involve sending prompts & data to companies outside of New Zealand. How comfortable are you with government information & your citizens' private information being sent to a third party, in another country?

Fortunately there are New Zealand companies like Catalyst Cloud that offer New Zealand hosted GPUs, capable of running AI models with-in our borders. However, this approach will limit the models you can choose as the top tier models (Claude, Gemini, ChatGPT) are hosted within their own offshore infrastructure.

Issue 2: Cultural Understanding

Saying that Large Language Models (LLM) are trained on large amounts of language doesn’t do the concept of “large” justice. AI companies have scraped every piece of content they can find to train & build their models.

Because of population demographics & online usage this means the bulk of the training data comes from the USA, a country which is different from New Zealand. From simple things like seasons & measurements (“Fall” happens in September, a “fluid ounce” is a measure of volume) through to the US education system (e.g. elementary schools, SAT exams, etc.). Even semantics become problematic - in New Zealand high schools are often called “colleges”, whereas in the US a “college” is a university - so “dropping out of college” has a very different meaning, depending on the country.

This can be countered with prompts, instructions & large amounts of your own data. Which brings us too…

Data Collection

Something AI is good at is summarising & inferring from large amounts of data. But data is king, you need a wide & deep dataset, or the AI models won’t have enough to work from.

We’ve already identified dropping out of high school means different things for different people. The only way we’re going to understand why students are dropping out of high school is to ask them.

So, we could get them to fill out a form, but forms tend to be clunky & lack nuance. But, more to the point - how many ex-high school students are going to fill out a form?

It’s possible to use AI agents which can phone people & carry out an interview with them using a remarkably human-like synthesised voice. But, AI chatbots have talked people into suicide & AI-induced psychosis, so should we really let them anywhere near potentially vulnerable people?

Another factor is that a form, or an AI-powered interview both assume we have contact information for people who’ve dropped out of high school. Ex-students who are the hardest to contact have important stories to tell & without their input our dataset is going to have gaping holes.

If we’re going to do this right we need teams of social workers who can talk to teachers, family & the wider community so they can locate the ex-students & interview them. Fortunately we can use AI to create transcripts from these interviews, which is a major time saver.

Data Analysis

If all goes well with our data collection we now have hundreds, possibly thousands of interview transcripts with ex-students who’ve dropped out of high school. This is where AI can really help.

AI can be used to identify trends, issues & opportunities from thousands of pages of text, in a fraction of the time it would take a human researcher.

AI can also be used to write reports & identify actions that can be taken to address factors they are causing students to drop-out. 

Human researchers are still needed, to guide the AI through data analysis. AI is fast & enthusiastic, but without guidance it will quickly go off the rails. No matter how much you instruct & prompt the AI to work from a New Zealand perspective the weight of its training will still lean towards the USA.

Why Can’t We Just Shortcut?

“Wait a minute” I hear you say “Surely heaps of research has already been done on why people drop out of high school? Why can’t we just use that?”

Which is a perfectly valid question. Let’s use Claude Sonnet 4.6 - which is one of the most powerful consumer models & the best at identifying bullsh#t - to ask:

“Using existing research can you please create a bullet point list of the most cost effective ways to prevent New Zealand students from dropping out of high school & to improve student (or ex-student) life outcomes?”

I’m going to also give Claude a copy of this article - up until this sentence - as context is key.

Claude’s Response:

Before jumping into the recommendations, a few honest caveats worth including in your article: existing research on school dropout in New Zealand is genuinely somewhat limited compared to the US corpus, which means any AI drawing on published literature will be pulling heavily from international studies. There's also a meaningful distinction between preventing dropout (structural/systemic interventions) & improving outcomes for those who've already left. The two lists overlap but aren't identical.

Here's what the evidence base — primarily from New Zealand's own ERO reports, Ministry of Education data, & international research adapted for NZ context — actually supports:

---

Preventing dropout (cost-effectiveness ranked roughly high to low)

• Early identification systems — Schools using attendance & engagement data to flag at-risk students early (Year 9–10) consistently show strong returns. NZ's Student Achievement Function Practitioners model demonstrated this. Catching disengagement before it becomes departure is far cheaper than any intervention after.

• Relationships before programmes — The single most replicated finding across NZ & international research: a consistent, trusted adult at school. Not a programme, not a resource — a person. Tuakana-tēina (peer mentoring) structures & whānau-connected pastoral care show particularly strong outcomes in NZ contexts.

• Attendance before anything else — Chronic absenteeism is the strongest predictor of dropout. Interventions targeting attendance (including transport barriers, housing instability, & school safety) outperform curriculum interventions in cost-effectiveness.

• Alternative education pathways — NZ's Gateway programme (vocational pathways from Year 11) & trades academies show solid evidence for retaining students who've disengaged from academic streams. Flexible timetabling & recognition of learning outside school reduce the binary "in or out" dynamic.

• Addressing underlying material hardship — Research consistently shows that food insecurity, housing instability, & poverty are primary drivers. Ka Ora, Ka Ako (school lunch programme) shows measurable attendance improvements at very low marginal cost.

• Mental health & learning support early — Unidentified learning differences (dyslexia, ADHD) & untreated mental health issues are consistently over-represented in dropout populations. Earlier screening & support reduces downstream costs substantially.

---

Improving outcomes for those who've already left

• Second-chance pathways (Trades, apprenticeships, Youthtown, Te Kura distance learning) — re-engagement is achievable when the barrier to return is low & the pathway is practical.

• NEET reduction programmes — Targeted outreach to 16–24 year olds not in employment, education, or training shows strong ROI. MSD's Youth Service contracts have evidence behind them, though implementation quality varies.

• Avoiding criminalisation — Research is unambiguous: contact with the justice system after dropout dramatically worsens life outcomes. Youth justice diversion programmes show strong cost-benefit ratios.

---

What the research can't reliably tell you

This is the part worth including in your article: most published research reflects programmes delivered in specific contexts (often low-decile urban schools, often Māori & Pasifika students, often with dedicated resourcing). The why varies enormously by region, ethnicity, family circumstance, & school culture — which is exactly your point about needing original data collection rather than assuming existing research transfers cleanly.

The AI hasn't fabricated any of the above, but it has compressed & generalised. Your social workers talking to actual ex-students in actual communities will find things this list doesn't anticipate.

To Conclude

At first reading Claude's response seems on target. But with a more critical eye - it gets most things right, misses some factors entirely, gets muddled in places & over-simplifies a complex issue (although I did ask for bullet points, so - my bad). Taking AI generated advice as gospel is fraught with potential issues & you're going to get higher quality advice via experienced humans using AI tools.

An important thing to keep in mind is that the advice wasn’t created by Claude - it was distilled from existing research created by real people. Without ongoing research feeding into the knowledge base it’s possible we’ll miss emerging issues - like AI chatbots telling students to drop out of high school.

AI is good at inference, summarising & writing in plain language - but it can’t collect the quantitative data that’s needed to generate meaningful & accurate responses. It can also fall victim to "data gaps" - where missing data can lead to uninformed or unhelpful responses.

Humans at the front line have a far better understanding of what data is important, what is data missing & what data is hard to source - whereas AI generally won't identify gaps & just work with what it's given.

Politics vs Reality

But, we need to address the elephant in the room & it’s hard to do this without “getting political”.

As Claude has said we need more teachers, social workers, support workers & support programmes. We need to invest in our young people for them to have fulfilling lives & structural issues like poverty are creating harmful outcomes. This has been shown time & time again & is absolutely beyond question.

There is certainly scope for AI to reduce office work, allowing people more time to work directly with those in need. But it still comes back to a need for more humans, working with humans, to help fix human problems at a human level.

The same Government who wants to cut public servant numbers & increase AI usage has already made significant cuts that go directly against Claude’s advice.

Is the desire to use AI based on productivity gains, or is it about having more control over the advice? It’s easy to think of a situation where someone is feeding their own political biases into AI in an attempt to get a response that fits their agenda.

So maybe the question is not so much “Should the Government use AI?”, but more “Is the Government willing to follow its recommendations?”

Banner Photo by Taylor Flowe