AI voice assistant case study, in development
Building an AI voice assistant for New Zealand’s lifestyle blocks.
How Changeable is working with lifestyle block owners Mark and Leigh Gray to build a capture-first AI app for the roughly 199,000 lifestyle blocks across New Zealand, where daily farm and animal care is currently tracked on paper, in someone’s memory, or not at all.
Live engagement. This case study describes work currently in progress. It will be updated as the app moves through its remaining build gates.
Project overview
An idea, tested twice, before the right one stood up.
Mark and Leigh Gray came to Changeable with an idea for an app to help lifestyle block owners manage their day-to-day. Rather than start building immediately, Changeable ran the idea through a structured use case assessment first: two early concepts, a general task app and a health-and-safety overlay, were tested and found not viable as standalone products.
The genuine opportunity turned out to be narrower and sharper: New Zealand has roughly 199,000 lifestyle blocks, and no AI-first voice app built specifically for the people running them. That gap became the brief. The app lets an owner speak or photograph what they see on their block, an animal, a fence line, a task just completed, and have AI turn it into structured, useful records without a form in sight.
Use case tested before building
Two earlier concepts were assessed and set aside before the real opportunity was confirmed.
Capture-first, voice-first
Speak or photograph what you see. AI structures it, not the other way around.
Client-owned from day one
The intellectual property and codebase sit with the client, not with Changeable, from the first line of code.
Staged, gated validation
The build proceeds through defined technical, quality and demand gates rather than a single go-live date.
The problem
Lifestyle blocks run on memory, paper and good intentions.
A lifestyle block sits somewhere between a hobby farm and a small commercial operation. Owners are looking after animals, paddocks, fencing, water systems and the odd emergency, often around a full-time job elsewhere. Very little of that daily work is written down anywhere useful.
Existing farm management software is built for commercial-scale operations, with the complexity and price to match. Lifestyle block owners are left with paper notebooks, memory and a search engine when something goes wrong with an animal or a paddock.
What the market gap looked like
- Roughly 199,000 lifestyle blocks across New Zealand, a genuinely large but underserved market
- No AI-first, voice-first product built specifically for this segment
- Existing farm software priced and designed for larger commercial operations
- Health and safety and animal welfare obligations that are easy to lose track of informally
- A meaningful retiree and semi-retired ownership segment, roughly a fifth to a quarter of the market, for whom simplicity and accessibility matter as much as capability
Architecture decisions
Seven non-negotiable decisions, locked before the build began
Before writing production code, Changeable and the client agreed a small set of architectural principles the build would not compromise on, whatever pressure came later.
Offline-first, two-stage commit
A lifestyle block is not always within mobile signal. Captures are saved locally first, then synced when connectivity returns.
Event sourcing for state changes
Every change to an animal, paddock or task record is stored as an event, not just overwritten, preserving a genuine history.
Token-conservation architecture
AI usage is designed to be economical by default, keeping running costs predictable as the user base grows.
Strict-liability suppression for high-risk advice
The AI is deliberately restricted from giving veterinary, structural, chemical or electrical advice. This boundary is hard-coded, not left to a prompt.
Trans-Tasman compliance split at schema level
New Zealand and Australian regulatory differences are built into the data model itself, ready for future expansion.
On-device audio preprocessing
Voice capture is cleaned up on the device before it is sent anywhere, improving accuracy in real paddock conditions.
Media-verification flywheel
Task completion is confirmed with a close-out photo rather than a self-reported checkbox, building a genuine evidence trail over time.
How the build is validated
Three gates, not one launch date
Rather than build everything and hope it works, the programme is structured around three specific proof points, each one a genuine go or no-go decision.
Offline sync works reliably
Including conflict resolution, so two changes made while offline resolve sensibly once the device reconnects.
The AI is useful and safe
Measured against a labelled set of real photographs, with a defined minimum accuracy threshold and zero tolerance for unsafe outputs. This is the genuine go or no-go point for the product.
Real people keep using it
A beta cohort is tracked for weekly active use and stated intent to pay, testing genuine demand rather than assumed demand.
What the app does
An irreducible core, with everything else able to flex.
Under pressure, features descope in a defined order, starting with insurance and compliance exports and ending, only as a last resort, with paddock treatment history. A small set of capabilities is treated as the non-negotiable heart of the product.
Questions
Questions about this engagement?
Common questions about how Changeable approaches an AI product build for an external client.
Why did the first two product ideas get rejected?
A general task app and a health-and-safety overlay were both assessed as part of early use case work and found not to be viable standalone products. Testing ideas honestly before building is part of how Changeable approaches every AI engagement, including its own.
Who owns the app and the code?
Mark Gray owns the intellectual property. The codebase sits in a client-owned repository from the very first commit, not inside Changeable’s own infrastructure.
Why restrict what the AI can advise on?
Advice on veterinary, structural, chemical or electrical matters carries real risk if it is wrong. That boundary is hard-coded into the system rather than left to careful prompting, which is a deliberate, permanent design decision rather than a temporary limitation.
Why build for offline use first?
Lifestyle blocks are not always within reliable mobile coverage. An app that only works with a live connection would fail at the exact moment it is needed, out in a paddock. Offline-first was treated as a non-negotiable architectural decision from day one.
What happens if a build gate is not met?
Each of the three gates, technical, AI quality and demand, is a genuine decision point. Features are designed to descope in a defined order under pressure, protecting the irreducible core rather than compromising it.
Is the app live yet?
The app is in active build, progressing through its defined validation gates. This case study will be updated as the engagement moves toward wider release.
Can Changeable run a similar use case process for our idea?
Yes. This is the same structured use case development process Changeable applies to every AI product engagement, testing an idea honestly before committing to a build.
Have an idea that needs testing before it needs building?
Changeable can run the same structured use case assessment on your concept, so you know whether it is worth building before a single line of code is written.