Halter's Mav: lessons for NZ firms putting AI over their own data

By DataForge team21 November 20265 min read

Halter has put an AI assistant over its collar data for 1,200 NZ dairy farms, and the launch shows other NZ firms what to check before doing the same.

What Halter launched

On 19 November Waikato agritech company Halter launched Mav, an AI assistant that lets dairy farmers ask plain-language questions about their herd and pasture. RNZ reported that farmers can ask how the herd looks that morning, which cows need attention, or how this season's pasture growth compares with earlier seasons. The answers come from data Halter's smart collars already collect, covering grazing history, animal behaviour, health and mating.

Farmers on Halter's Pro and Unlimited plans can already use Mav, which puts it in the hands of 1,200 dairy customers across New Zealand. A launch overseas is on the way, though Halter hasn't given any dates yet. Rural News reports that farmers asked more than 100,000 questions during the beta. Farmers can also schedule a question so the answer arrives as a daily update. According to Halter, the model learns from pooled records with identifying details stripped out, spanning over half a billion collar days, 1.2 million health cases, 7 million heat detections and 10,000 breeding seasons in total.

Halter also opened what it calls an Agricultural Intelligence Archive in a converted cattle underpass at its Morrinsville research farm. The archive is a physical record of verified farming insights, inspired by seed banks. Farmers Weekly says the company had considered building an assistant like Mav for years and committed fully to the product over the past year.

Why the launch matters beyond dairy

Most AI adoption coverage in New Zealand concerns office productivity tools: chat assistants, meeting summaries, drafting help. Mav belongs to a different category. A company has spent years collecting structured operational data, and it has put a language interface over that data so frontline staff can ask questions without opening a dashboard or exporting a spreadsheet. For many NZ businesses the operational data already exists in a warehouse management system, an ERP, a job management tool or a fleet telematics feed. Getting answers out of that data still depends on a few people who know which report to run.

The time saving in RNZ's story is easy to picture in other settings. Andrew Macky, who runs a 300-cow Waikato farm, told RNZ that sorting cows into mobs with Mav takes about 30 seconds, where it might otherwise take 10 to 15 minutes. A dispatcher grouping jobs by suburb works the same way, and so does a warehouse lead deciding which orders to pick first. In each case the person already knows what they want and loses time assembling the data. Rural News gives a sharper example from mating season. On a large farm around 50 cows can come into heat on the same day, and Mav weighs heat history, timing and intensity for every cow at once. That kind of many-variable sorting is where a well-grounded assistant earns its keep.

Mav's apparent success depends on Halter's data foundation. Halter's collars have generated more than 1.2 trillion data records, according to Rural News, and every record comes from the same hardware in a consistent format. Few mid-sized NZ firms have data that clean. Customer records get duplicated, job codes get used inconsistently, and spreadsheets sit beside the core system. An assistant laid over messy data will answer with confidence and still be wrong, and staff will stop trusting it after a few bad answers.

The questions Halter's launch leaves open

Halter's founder Craig Piggott told RNZ the company is "in the business of selling products and collars, not selling data." The coverage we've seen doesn't describe the consent terms behind the aggregated training data, how farmers can opt out, or how Mav's answers are checked for accuracy. Farmers Weekly says only that Halter works to make sure farmers can trust Mav's advice and keeps improving the assistant. None of the reports we've seen state an error rate or describe a validation method.

Those gaps aren't unusual for a product launch, but the same questions apply to any business buying a vendor's AI feature or building its own. Halter's model pools anonymised data from many customers to train an assistant that serves each customer individually. Plenty of NZ software vendors are moving toward the same design, from accounting platforms to field service tools. If your data goes into a vendor's shared model, your contract needs to say what's pooled, how it's anonymised, and whether you can withdraw. Under the Privacy Act 2020, aggregated data about animals carries little risk. Aggregated data about customers or staff needs much closer scrutiny.

The verified archive is an interesting design choice. Halter records Mav's findings only after checking them, which separates answers the company stands behind from answers the model generated on the fly. Businesses can copy the principle without building a hillside vault. An internal register of AI-produced findings that a person has reviewed and approved gives staff a source to rely on, and it gives auditors a trail to follow.

What NZ businesses should do next

Halter's launch is a useful template for firms that hold years of operational data and want frontline staff to use it more. These steps apply whether the assistant comes from a vendor or gets built in-house:

  • Find the question staff ask most. Pick a decision that takes 10 minutes of report-pulling today, such as job grouping, stock reordering or overdue maintenance, and measure how long it takes now so you can compare later.
  • Audit the data under that question before choosing a tool. Check for duplicates, inconsistent codes and gaps. Halter's assistant works because its collar data is uniform, and yours needs to reach a similar standard for the chosen use case.
  • Read the vendor's data terms for embedded AI features. Ask whether your data trains a shared model, how anonymisation works, and how you can opt out. Get the answers in writing.
  • Decide where the model runs. If the data is commercially sensitive or covers people, consider whether a private AI server suits you better than a shared cloud model.
  • Set up a review step for findings people will act on. Keep a simple register of AI outputs that a named person has checked, and treat unreviewed answers as drafts.
  • Run a limited beta with real users. Halter collected more than 100,000 questions before the general release. A few weeks with a small team will show you which questions people actually ask and where the answers go wrong.

Firms that want help with the data audit, vendor questions and governance around a first operational AI assistant can find out more about our AI consultancy work.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS