Research note NZ site feasibility Source-backed
The research question

How can AI strengthen the judgement a person brings to a site?

Everything below is about one thing: making the judgement a professional brings to a piece of land sharper and better informed, without taking the call away from them. The site is already described in public records. The work is putting all of it in front of the person before they decide. This note is grounded in New Zealand public data and in a prototype we have built.

01The judgementWhat a person is reading for

Before any line is drawn, a person is already judging the land.

The first hour on a site is not drawing. It is judgement. A professional is weighing a small number of questions that shape everything downstream, and getting them wrong is expensive to undo once design has hardened. In New Zealand that judgement also has a signature attached: restricted building work must be done or supervised by a named, individually accountable person, and a producer statement is recorded as a professional opinion, not a guarantee.45

What the first judgement actually weighs
01Buildability

Can this land carry the building the client has in mind, and at what difficulty and cost.

02Constraint load

How far the zone, overlays and boundaries will shape, limit or rule out the design.

03Risk exposure

Flood, slope, ground and other hazards that change cost, consent and who carries the liability.

04Brief fit

Whether the client's brief is realistic on this particular piece of land, or needs to flex.

05Cost of finding out

How much time and money it will take to remove the unknowns that remain.

02The stakesWhy the first hour carries weight

The early call sets the cost, and the cost of getting it wrong is large.

Early decisions have the most leverage over a project and are the cheapest to change, which is the principle behind the long-cited MacLeamy curve. The flip side is that an early misread is the most expensive thing to undo, and New Zealand has paid for that at scale. Build-quality failures are estimated to cost about a tenth of the residential construction sector's value, and the bill for defective buildings has repeatedly landed on councils and ratepayers rather than the parties who made the error.23

The cost of getting it wrong, in New Zealand
~$2.5b

Estimated annual gain to the economy from fixing build-quality failures, around 10% of the residential construction sector's value lost to defects.

BRANZ / NZIER · 2026
92%

Of new houses surveyed carried at least one compliance defect, a signal of how often things are missed before and during the build.

BRANZ survey · 2014
~$1b

Paid by councils and ratepayers over a decade for defective buildings under joint-and-several liability, about a third of it for parties who avoided their share.

MBIE / Sapere · 2008–2018

This is the backdrop the leaky-building crisis set, an estimated repair bill of around $11 billion across some 42,000 buildings. It is also why the country is now moving from joint-and-several to proportionate liability and requiring designers to carry professional indemnity insurance.3 A site judgement that is better informed, and that leaves a defensible record of what was known, sits squarely inside that shift.

03The problemWhy the call is made half-blind

The information exists. The person rarely sees all of it in time.

Almost everything needed to judge a site is public. It is also scattered: parcels and terrain sit with LINZ, planning rules sit with each council, and hazards are spread across national models and regional registers, in formats that do not line up. There is no single national source that returns a property's full planning and hazard picture, so in practice the early judgement gets made on whatever could be pulled together quickly.9

The gap

A site is fully described in public records. The person almost never sees the whole description before they have to decide.

The limit is not knowledge. It is assembly: the time and effort to bring scattered public records into one place a person can actually read. That cost is what quietly narrows the early judgement.

04What AI assemblesThe whole picture, in one place

AI can lay out the whole site before the judgement is made.

The first thing AI changes is reach. Much of the base is genuinely open and machine-readable: LINZ publishes parcels and titles under a Creative Commons licence with a public API, and national LiDAR now covers more than 80% of the country as a one-metre terrain model, free to use.68 We have built a working version of this layer, which is what grounds the rest of this note (see the NZ site intelligence case study). The honest part is that the layers are not equal in quality, so the assembled picture has to carry that distinction on its face.

The assembled layer, with source and reliability on every line
LayerSource and licenceCoverage and freshnessReliability
Parcels and titlesLINZ Data Service, CC‑BY 4.0National; WFS and API access; titles updated weekly. Owner names sit behind a separate licence.Open and current
Legal boundary precisionLINZ digital cadastre±0.2 m in survey-accurate urban areas, widening to ±5 m and up to ±100 m in non-survey-accurate rural land.Varies by location
Terrain and slopeLINZ / NZ Elevation, on AWS1 m DEM and DSM; over 80% of NZ; each survey carries a capture date and can predate recent earthworks.Open where flown
Planning zones and overlaysCouncil GIS, per districtStandardised zones under the National Planning Standards, but no national dataset; assembled council by council, with proposed and operative plans coexisting.Fragmented
Flood hazardEarth Sciences NZ; councilsFirst national model covers 256 flood plains at a 1% annual chance, but it is regional and modelled, not a property-level verdict, and excludes sea-level rise.Modelled, indicative
LiquefactionGNS / MBIE Level A–BRegional susceptibility domains from desktop assessment; consent-grade decisions need site-specific Level C–D investigation.Desktop only
LandslideGNS Landslide DatabaseOver 100,000 recorded events; an inventory of where landslides have happened, not a prediction of where they will.Past events only
Open and current

Authoritative, machine-readable, and safe to lean on.

Varies / fragmented

Accurate in places, uneven elsewhere; check before relying.

Modelled or historic

Indicative only; not a substitute for site investigation.

05MethodHow the picture gets built

Two ways to get there: general research, or a tool of your own.

There are two ways a firm actually produces this layer today, and they are easy to confuse. One is to point a general deep-research model at the open web. The other is to build a tool that reads your own authoritative sources directly. The evidence is now clear that they fail and succeed in opposite places, which is the whole reason to be deliberate about which does what.14

Why the split matters, from the published evidence
>60%

Of source-identification queries answered wrong by leading AI search tools, which tended to be confident rather than decline.

Tow Center, CJR · 2025
~97%

Of statements in one deep-research system's answers were not supported by its own cited sources, at the high end of an audit across systems.

DeepTRACE audit · 2025
~51%

Best score a leading model reached on multi-step geospatial tasks, with documented errors in basic geometry.

GeoBenchX · 2025
Approach A

General deep research

Point a capable model, the ChatGPT or Claude deep-research style, at the open web and ask it to research the site and write up what it finds.

Strong at
  • Open-ended questions with no single fixed source
  • Local context, precedent and neighbourhood character
  • Speed, with nothing to build or maintain
  • Pulling a readable narrative together quickly
Weak at
  • Authoritative geospatial and planning data
  • Computing slope or geometry from terrain
  • Provenance, and citing sources that hold up
  • Returning the same answer twice
Approach B

A tool over your own data

Wire directly into the authoritative sources, LINZ, council and hazard registers, and hold them as structured data you control.

Strong at
  • Authoritative, current facts with the source attached
  • Structured terrain and geometry you can measure on
  • Freshness and provenance on every field
  • Repeatable results that compound across projects
Weak at
  • The open-ended questions outside its sources
  • Build and maintenance cost
  • Breadth: only as wide as what you wire in
  • Flexing to a question you did not design for

The reason to ground the load-bearing facts in your own data is not a preference, it is measured. Grounding generation in a verified source has been shown to cut hallucinated steps in structured output from around a fifth to under one in thirteen, while general web agents are documented as non-deterministic between runs even on identical prompts.161819

The good output

The strongest result is not one approach or the other. It is the custom tool holding the facts a judgement must not get wrong, with general research working the edges on top of it.

The spine

Parcel, terrain, planning and hazard come from the structured tool, with source and date attached, so the load-bearing facts are firm.

The edges

General research handles precedent, local character and the questions with no single source, layered over that spine rather than inventing it.

The call

Both feed one readable view, and the professional still makes the judgement, now on firm ground rather than a guessed-at picture.

06The amplificationWhere the call gets sharper

Assembly is the start. The judgement gets stronger in four ways.

Putting everything in one place is only the first gain. The larger one is what a person can now see: the conflicts, gaps and interactions that stay invisible while the data is scattered, and that are exactly where judgement earns its keep.

Where AI value already sits

Architects put AI's value earliest in the work, exactly where site judgement happens.

In a 2026 survey of around 800 architects and designers, 43% named concept and pre-design as the stage where AI adds the most value, ahead of documentation and delivery.1

Nothing judged blind

Every factor that bears on the site is in front of the person, so no part of the call rests on a guess about what was never checked.

Conflicts surfaced

Where two records disagree, the disagreement is shown rather than silently resolved, which is precisely where a person's judgement is needed.

Missing data made loud

What is not known is shown as not known, so the person can weigh the unknown instead of mistaking an absence of data for an absence of risk.

The question can move

A person can ask what if the access sits here, or the platform there, against the whole site in seconds, and judge more options before committing to one.

Where the person stays

The call is still theirs. They make it with the whole site in front of them, faster, and with the unknowns named, instead of on a fraction of the picture.

07The riskWhere AI would weaken the call

A site layer that looks finished is the real danger.

The failure mode is not a wrong number. It is false confidence. A clean, assembled site layer looks authoritative, and the temptation is to stop checking. Much of the public record is indicative, modelled or out of date, and a layer that hides that strengthens nothing.

The trap
Treat the assembled layer as the truth.

Cadastral boundaries can be metres out in rural areas, LiDAR can predate recent earthworks, planning layers vary between proposed and operative, and the national flood model is explicitly regional and indicative rather than a property-level assessment. The absence of a flood note on a LIM does not mean a property will not flood, and liquefaction maps are desktop-grade until a site investigation is done.7101112 A layer that smooths over this reads as more certain than it is, and quietly weakens the judgement it was meant to strengthen.

The discipline
Make every fact carry its source and its age.

Each assembled fact shows where it came from, how fresh it is, and whether it is open record or a model. Missing data stays loud. Trust stays calibrated, so the person leans on the firm parts and holds judgement on the soft ones, which is what keeps the call sound and the record defensible.

08What changesThe stronger first hour

A better-informed first hour changes the whole project.

When the early judgement is made on the full picture, the cost of getting it wrong later falls. The timing is also right: the country's planning law is being rewritten with natural-hazard risk as an explicit objective, and liability is moving to a proportionate model backed by mandatory professional indemnity cover.320 Both reward a site assessment that is better informed and that leaves a clear record of what was known and from where.

What it changes

Fewer late surprises

Constraints that usually appear during consent are visible in the first hour instead.

Faster feasibility

A defensible go or no-go view forms in a sitting, not over a week of chasing records.

Cleaner brief conversations

The client hears what the land allows early, while the brief can still flex around it.

A reviewable trail

Every site fact carries its source, so the early judgement can be defended later.

Where to start

Assemble, do not decide

Use AI to gather and lay out the records, and keep the call with the person.

Show the seams

Surface conflicts and missing data first, because that is where judgement earns its keep.

Keep the sources visible

Every fact should show where it came from, how old it is, and whether it is record or model.

Start with the hard sites

The constrained, hazard-heavy ones, where a fuller picture changes the most.

Closing thought

The goal is not a site that reads itself.

It is a person who can read the whole site at once. AI does the gathering, the cross-checking and the surfacing, so the judgement a professional brings to the land is made on everything that is known, not a fraction of it. The site still gets judged by a person. They just stop judging it half-blind.

Matt Strawbridge · Landform ResearchAotearoa New Zealand · 2026
References and sources
  1. Chaos & Architizer. State of AI in Architecture survey (~800 respondents). 2026. blog.chaos.com
  2. BRANZ / NZIER. The economic cost of building quality defects (ER49); incl. 2014 defects survey. 2026, reported via RNZ and The Conversation. rnz.co.nz
  3. MBIE / New Zealand Government. Building system reform: consumer protections factsheet (proportionate liability, PI insurance, council payouts). 2025. beehive.govt.nz
  4. Licensed Building Practitioners (MBIE). Restricted building work and the design process. 2026. lbp.govt.nz
  5. Building Performance (MBIE). Producer statements. 2026. building.govt.nz
  6. Toitū Te Whenua LINZ. LINZ Data Service: licensing and using data (CC‑BY 4.0). 2025. linz.govt.nz
  7. Toitū Te Whenua LINZ. Accuracy of the digital cadastre. 2025. linz.govt.nz
  8. AWS Registry of Open Data / LINZ. NZ Elevation: 1 m DEM and DSM, CC‑BY 4.0. 2025. registry.opendata.aws
  9. Ministry for the Environment. National Planning Standards. 2019, updated 2022. environment.govt.nz
  10. Earth Sciences New Zealand (NIWA). Nationwide study reveals escalating flood risk (first national flood model). 2025. earthsciences.nz
  11. BRANZ (Level). Site analysis: flood risk, and the limits of LIM hazard data. level.org.nz
  12. Building Performance (MBIE) / NZGS. Regional liquefaction vulnerability assessment (Levels A–D). 2017–2021. building.govt.nz
  13. GNS Science / Earth Sciences NZ. New Zealand Landslide Database. 2017–present. gns.cri.nz
  14. OpenAI. Introducing deep research. 2025. openai.com
  15. Tow Center for Digital Journalism, Columbia Journalism Review. AI search has a citation problem. 2025. cjr.org
  16. Venkit et al. DeepTRACE: auditing deep research AI systems. arXiv:2509.04499. 2025. arxiv.org
  17. Krechetova & Kochedykov. GeoBenchX: benchmarking LLMs on multistep geospatial tasks. arXiv:2503.18129. 2025. arxiv.org
  18. Béchard & Ayala. Reducing hallucination in structured outputs via retrieval-augmented generation. NAACL 2024. aclanthology.org
  19. Anthropic. How we built our multi-agent research system. 2025. anthropic.com
  20. Ministry for the Environment. Resource Management reform: Planning Bill and Natural Environment Bill. 2025. environment.govt.nz
← Back to Landform Research