Built for how one nursery works. Productised for the industry.

The business ran on one person’s memory

John had searched for nursery software and found nothing that fit. What he had instead was a spreadsheet, paper labels, and everything important held in his own head.

That works until it doesn’t. If he’s sick or away, nobody else can answer what’s ready, what it cost, or where it is.

Plants die, prices are wrong, and money gets lost.

A nursery is more complicated than it looks

Hundreds of thousands of plants. Dozens to hundreds of species, split across hundreds of batches, each at a different lifecycle stage, spread across propagation houses, growing bays and outdoor beds.

And orders arriving for delivery anywhere from next week to next year, each one drawing from different groups of plants that don’t exist yet.

That combination is why spreadsheets fail here. It isn’t volume, it’s that everything is changing state independently, all the time.

We’re an AI company. We started with computer vision, and moved away from it.

The original thinking was AI-first: point a phone at a bench, get a plant count.

But the main problem isn’t counting stock. It’s knowing what’s happening and when. Computer vision would have solved a useful but small part of that, and it’s genuinely hard in a nursery: hundreds of species, plants overlapping, faded tags, half the stock under shade cloth.

Seeing all the data and making informed decisions is what actually counts.

So the goal changed: build a seed-to-sale platform for managing nursery inventory.

The risks we wrote down before building

The bet: inventory and sales alone would be enough

We scoped hard. No planning module, no reporting, no forecasting. Get the core functions right first: batches, inventory, locations, quotes, orders, invoices.

Two surfaces, not one: an on-site app for the people working the benches, and an admin view for the person running the business. The task system was phased so later stages wouldn’t need a destructive migration.

Building it, and finding where AI actually belonged

I ran moderated testing with nursery staff, six tasks each. Two of the six failed for every participant, both in the same part of the product, which told us we’d never properly defined what belonged where.

The most useful finding wasn’t a fix, it was a reframe. One tester explained that his quotes already exist. They arrive as emails with thirty species on them. His job is to check that list against stock and price it, not retype it.

Review instead of constructing.

That became the AI feature we did ship: forward the customer’s email, get a drafted quote to check. It only works because the inventory layer underneath is trustworthy enough to check against.

Going to market through a small industry

New Zealand nurseries are a small, connected market. Word of mouth and relationships decide things, so early traction experiments pointed at trade shows, in-person selling and referral rather than advertising.

We had strong trust with our pilot customer, and he already had a stall booked at the Hamilton Trade Days. So that became our public launch.

Four assumptions, written down before we went

We assumed	What came back
$300/month is the right price	Under priced
Nurseries are disappointed with existing suppliers	Confirmed
The conceptual model is right: batches, inventory, sales	Confirmed. Lifecycles, batches and the data on them, the sales model, and the split between the on-site app and the admin view all transferred. Only details needed changing
The system can flex where it needs to	Partly. It needed specific changes in the details

Outcomes

Early stage. This is a case study about decision quality under uncertainty, not proven commercial outcomes.