NZ’s first property MCP connector for Claude and ChatGPT

When a New Zealand agent asks Claude or ChatGPT a property question — comps in a suburb, a market narrative, a follow-up after a CMA — the model is only as good as the data behind it.

Generic chat tools do not know today’s REINZ unconditional sales. They invent fluency. You still own the risk under REA’s Gen AI guidance and your appraisal duties under rule 10.2.

Newton is built for that gap: a New Zealand property research assistant that can connect the AI tools you already use to live property evidence, with sources you can open before anything client-facing leaves the office.

This piece explains the connector idea in plain English for agents and agency leaders. It is product journalism, not a legal opinion — and it does not replace REA or Privacy Act duties. It also deliberately avoids marketing slogans that belong on LinkedIn rather than in a compliance-aware hub post.

Claude and ChatGPT are good at the work you already do in chat: drafting, questioning, comparing options, stress-testing a story, turning rough notes into a clearer email, helping a junior organise thoughts before a listing presentation.

They are not a substitute for verified NZ sales evidence. Without a property data connection, “AI for real estate” is just confident prose — and REA’s guidance is clear that polished, wrong output is still your problem. Accuracy, relevance and completeness remain human duties. Gen AI error is not a defence.

So the healthy office mental model is three layers:

Miss any layer and you either get slow work, invented work, or unsupervised work. None of those are a strategy.

You do not need the jargon soup. Think of a connector as a plug that lets an AI assistant call a specific tool or data source instead of guessing from general training.

In practice for property teams:

The job of the connector is not “AI in general.” It is live property evidence in the assistant — so answers can be checked rather than invented. That is the entire point for listing and advisory work.

REA’s Gen AI guidance stresses accuracy, relevance and completeness — especially for appraisals and marketing. Rule 10.2 still requires a written appraisal supported by comparable sales that realistically reflect current conditions. An automated or AI-shaped mid-point is not that appraisal. Rule 10.3 still requires a written explanation when comps are thin.

A connector that surfaces checkable sales context helps the research step. It does not delete human oversight, privacy rules, or the need to label automated estimates honestly. REINZ Companion Terms are blunt that Estimate outputs are not appraisals and must not be presented as a CMA.

Privacy still applies: do not paste client LIMs, vendor files or personal information into consumer chats. Connectors and purpose-built property tools are about evidence access — not a free pass to upload confidential PDFs. REA and the Privacy Commissioner both push the same direction: personal and client information does not belong in uncontrolled external prompts.

That is the difference between “we use AI” and “we use AI on NZ property evidence that survives checking.”

Atlas drafts branded CMAs on live NZ data — useful when the deliverable is a keepable appraisal pack. Newton answers property follow-ups with sources shown — including through a connector in Claude or ChatGPT when your team works that way. Orbit sits with the wider PropertyLM stack for day-to-day team workflows.

None of them replace a licensee’s signature. They make the evidence pack harder to poke holes in — which is what vendors, lenders and Complaints Assessment Committees actually care about.

Explore Newton on PropertyLM.

A connector is not a magic appraisal machine. It is not permission to skip viewing. It is not a privacy bypass. It is not a substitute for agency policy. It is not an argument that every question should be answered inside a chat window.

Used well, it reduces invented fluency by putting checkable NZ property context next to the assistant. Used badly, it becomes another place people forget to open the source. Training still matters.

Answer those five and the connector becomes infrastructure. Skip them and it becomes another shiny risk.

Without a property connector, Monday often looks like this: someone asks a chat model for suburb comps, gets a fluent answer, half-trusts it, then spends the afternoon reconstructing reality from the sales database anyway — or worse, does not.

With a connector and a trained team, Monday looks more like this: ask the question in the assistant you already use; pull NZ property context through Newton; open the sources; draft; check; send. The assistant stays in the workflow. The evidence stops being optional.

That is the operational claim worth making in a hub post: fewer invented suburb stories, more checkable research inside the tools people already open — with REA-aligned human ownership still intact.

Day one: show the difference between a generic suburb answer and an answer with openable NZ sources. Day two: practise a bad path — pasting a redacted LIM — and stop it cold. Day three: practise labelling an automated estimate correctly in a vendor email. Day four: walk a junior through opening every source before they forward a chat summary. Day five: spot-check five outputs and coach in public on what “good” looked like.

That week costs less than one messy complaint file. It also makes the connector feel like infrastructure rather than a novelty — which is the only sustainable way to introduce property-aware assistants into a licensed office.

Keep the vendor sentence handy, keep rule 10.2 language accurate, and keep confidential files out of consumer prompts. The connector helps the evidence layer. It does not forgive skipping the accountability layer.

— PropertyLM.