A transparent NZ sector-calibrated policy sandbox for testing AI policy tradeoffs under uncertainty.
NZAIS is a whole-economy scenario model for comparing AI policy approaches in New Zealand.
It exists because New Zealand's current AI policy discussion rests on a fragmented baseline. Adoption figures in circulation range from 32% to 87%, depending on who was surveyed, what population was sampled, and how "adoption" was defined. No two sources measure the same thing. Policy built on any single number is standing on weak ground.
That means much of the public conversation around "NZ AI adoption" is flatter and more confident than the evidence supports.
This project replaces single-number thinking with a structured, sector-calibrated framework that compares different policy designs under uncertainty.
Does aggregate AI policy, designed around a single national average, systematically produce worse outcomes than sector-targeted policy in New Zealand?
A related question sits underneath it:
When is it better to support direct adoption in sectors, and when is it better to invest in enabling capacity - technology, skills, procurement, infrastructure, and diffusion?
- a policy sandbox - for comparing scenarios, not forecasting GDP
- a structured synthesis of fragmented NZ evidence
- a whole-economy model with tiered sector coverage
- transparent - assumptions are labelled, not hidden
- a forecasting engine
- a claim to exact future outcomes
- a definitive adoption ranking
- a black box
The evidence shows clear sector differences:
| Sector | Archetype | Why it matters |
|---|---|---|
| Agriculture | Targeted-policy beneficiary | NZ's most distinctive sector. No published adoption rate. Long diffusion timeline. |
| Manufacturing | Productivity sweet spot | 58% adoption. Clearest productivity case. Investment Boost lowers barriers. |
| Professional Services | Diminishing returns test | Highest adoption. Further investment yields smaller marginal gains. |
| Public Sector | Employment story | Lowest adoption, fastest acceleration. Capability, culture, and scale are the barriers. |
| Technology | Supply-side enabler | Not just an adopter - it enables the other sectors. $24B GDP, $11.4B exports. |
| Healthcare | Equity-constrained high-potential | 62% adoption. Admin AI vs clinical AI. Equity and governance are binding constraints. |
| Construction | Worst case for aggregate policy | No published adoption rate. 95% of firms < 10 people. Lowest digital maturity. |
| Financial Services | Regulatory throttle | High adoption constrained by dual regulators (FMA + RBNZ). |
| Retail & Wholesale | Consumer-facing displacement | Largest sector by weight. Wholesale 64% adoption. Retail = displacement story. |
These differences are not cosmetic. They change how policy works.
The model uses a tiered structure covering all 19 ANZSIC Level 1 sectors.
ANZSIC (Australian and New Zealand Standard Industrial Classification) is the official system used by Stats NZ to classify all economic activity into sectors. When this project says "all 19 ANZSIC Level 1 sectors," it means every part of the economy is represented - from agriculture to healthcare to retail. See GLOSSARY.md for all project terms.
Agriculture, Manufacturing, Professional Services, Public Sector, Technology, Healthcare, Construction, Financial Services, Retail & Wholesale
Education & Training, Transport & Warehousing, Accommodation & Food Services, Administrative & Support Services, Information Media & Telecommunications, Utilities
Mining, Rental/Hiring/Real Estate, Arts & Recreation, Other Services
Why tiered? Because a 9-sector-only model covers ~61% of GDP. That is not enough for an honest economy-wide aggregate-policy comparison. The whole economy must be represented.
Broad economy-wide allocation by GDP share or equivalent rule.
Support focused on sectors where adoption is lagging, bottlenecks are acute, or direct intervention unlocks productivity or public value.
Support focused on enabling capacity - technology, skills, infrastructure, procurement, and diffusion mechanisms that raise adoption capacity across the economy.
The central analytical tension is not only aggregate versus targeted. It is also:
direct adoption support versus enabling-system investment.
The project draws on 10 completed research analyses and 38 catalogued data sources:
| Source | Key contribution |
|---|---|
| OECD "Miracle or Myth?" (2024) | Productivity calibration: 0.25–0.6pp TFP growth from AI over 10 years |
| NZ Treasury AN 24/06 | NZ structural barriers and diffusion lag |
| AI Forum NZ (Sep 2024) | Adoption data, methodology critique, NZ case studies |
| Datacom State of AI (2024) | 66% adoption (100+ employee firms), governance gaps |
| PM's Chief Science Advisor (2023) | Healthcare: admin vs clinical AI, equity constraints |
| Public sector documents (4) | Cabinet approach, GCDO framework, "chilling effect", barriers |
| RBNZ "Rise of the Machines" (2025) | Financial stability risks, regulatory vacuum |
| FMA AI Research (2024) | Financial services adoption (13 respondents) |
| NZTech Manifesto (2026) | Tech sector: $24B GDP, 119,520 jobs, $11.4B exports |
| Stats NZ (multiple) | GDP, employment, business demography - all 19 sectors |
The source comparison - showing that adoption figures range from 32% to 87% with no consistent definition - is a foundational contribution of this project.
The primary reference architecture is adapted from Strauss (ai-web-economy-simulator) - a mechanism-design ODE model with:
- bounded state variables
- one-mechanism-per-layer causal structure
- named policy scenarios
- mathematical specification before code
The adaptation replaces the two-population platform/creator model with a 19-sector tiered economy where each sector block has interpretable state variables for adoption, absorptive capacity, productivity effect, and labour pressure.
Tech stack is not yet decided. The methods and scope come first.
- Project scope locked
- Whole-economy tiered structure defined
- Formal paper outline written
- Methods note drafted
- 10 research analyses completed
- 38 data sources catalogued
- 348-cell parameter table created
- 122/348 cells pre-populated (35%)
- Primary reference repo selected (Strauss)
- Architecture adaptation brief written
- State-variable specification drafted
- First-pass equations
- Tech stack decision
- Version 1 model build
- Paper draft
- External review
- Public-facing interactive layer
nzais/
├── README.md # This file
├── SCOPE.md # Formal project scope
├── PAPER-OUTLINE.md # Paper structure
├── METHODS.md # Methods note
├── STATE-VARIABLES.md # Critique-ready state variable specification
├── FAQ.md # Comprehensive FAQ
├── GLOSSARY.md # Plain-language glossary of project terms
├── SOURCES.md # Key source documents with URLs
├── CONTRIBUTING.md # Collaborator brief
├── docs/
│ ├── critical-analysis.md
│ ├── repo-selection.md
│ ├── architecture-adaptation.md
│ ├── provenance-analysis.md
│ └── data-source-catalogue.md
├── data/
│ ├── sector-parameter-table.csv # 348-cell parameter table
│ └── raw/ # Source data files
├── research/
│ └── *.md # Per-source analysis notes
└── src/ # Model code (TBD)
This project is currently private and in foundation phase.
If you are interested in contributing - as an economist, sector expert, policy analyst, or technically literate critic - the most useful next step is a short conversation focused on:
- what the project is missing
- what it is overclaiming
- what evidence or structure would make it genuinely more useful
See CONTRIBUTING.md for contribution modes and the collaborator brief.
This project is developed under AI for Good New Zealand - a community dedicated to ensuring AI benefits all New Zealanders.
This project is led by the AI for Good New Zealand collective - a community of practitioners, economists, sector specialists, and policy analysts working to ensure AI benefits all of Aotearoa.
MIT - see LICENSE.
