Assessing AI and Robotics Exposure in the New Zealand Labour Market Using Large Language Models

Key findings

• This study measures artificial intelligence (AI) and robotics exposure for New Zealand occupations using 10 large language models (LLMs) to assess automation potential.
• LLM-based exposure scores show strong agreement for robotics exposure but more variation for AI exposure, suggesting greater uncertainty in assessing cognitive task automation compared to physical task automation.
• AI exposure is most strongly associated with professional, managerial, and administrative occupations and with middle to upper-middle income brackets ($50,000 to $150,000). Machinery operators, labourers, and community workers show significantly lower AI exposure.
• Robotics exposure concentrates in machinery operator and labourer occupations. Workers with no qualifications are generally more exposed than those with secondary or tertiary qualifications.
• Joint exposure to both technologies is widespread across occupations. Approximately 30% of workers face high combined exposure, while less than 5% face low exposure, indicating that few occupations can avoid engagement with at least one of these technological shifts.

Why we did this research

AI and robotics represent distinct but related sources of technological change in the labour market. AI primarily affects cognitive and analytical tasks, while robotics targets physical and manual work. Understanding exposure to both technologies and their interaction matter for anticipating where and how jobs may change.

Central banks are interested in these dynamics because shifts in labour markets can affect wages and inflation. Early evidence suggests AI may disproportionately affect higher-skilled workers, potentially reversing historical patterns where technology primarily displaced lower-skilled work. Robotics exposure, by contrast, remains concentrated in manual and production roles.

This research develops New Zealand-specific measures of AI and robotics exposure, recognising that our distinct occupational mix may face different impacts from these technologies.

What data have we used?

The analysis combines occupation classification data, census employment statistics, and exposure assessments from multiple large language models. These are summarised in the table below.