AI in Agriculture

Executive Summary

• The sector now has a clearer adoption baseline, and it is lower than the commercial success stories suggest. AWS-commissioned research reported AI use at just under 40% across agriculture, forestry and fishing, compared with 50% across New Zealand businesses overall. Most sector use remains basic rather than deeply embedded.
• Commercial adoption is concentrated in targeted workflows. Halter, Aimer and Hectre are operating or scaling systems for livestock management, pasture measurement and fruit-quality assessment. These are more mature than general-purpose generative AI on farms.
• Virtual fencing has moved into structured national validation. The five-year LIFT programme has onboarded nine commercial sheep and beef farms, with collars deployed and production, financial and environmental baselines being established. It is now an evidence programme, not merely a funding announcement.
• Horticultural AI is progressing from grant-funded concept to laboratory and prototype work. Hectre has opened an Auckland research laboratory for hyperspectral internal-fruit analysis, but commercial validation is not expected until 2028.
• Independent evidence of returns remains thin. Early farm accounts are encouraging, but most reported benefits are vendor-, farmer- or programme-reported. There is still no national, independently verified estimate of AI’s productivity or environmental impact across New Zealand farms.
• Reliability, connectivity, skills and data governance are now the binding constraints. Hardware failures, weak rural connectivity, fragmented data and limited AI capability are proving as important as model performance.

The sector is therefore commercially active but not broadly transformed. New Zealand has globally competitive agritech companies, but adoption at the farm gate remains selective, uneven and highly dependent on whether a tool solves a specific operational problem.

What Has Changed Since the Last Update

National adoption evidence now points to a two-speed sector

The most important new development is better measurement of the gap between AI activity and AI adoption.

AWS’s September 2026 research found that 50% of New Zealand businesses were consistently using AI, up from 42% a year earlier. However, reporting on the same research by RNZ and other New Zealand media placed adoption across agriculture, forestry and fishing at just under 40%. Only about 10% of users in the primary-sector grouping were described as having reached advanced use. The figures are vendor-commissioned and cover agriculture, forestry and fishing rather than agriculture alone, so they should be treated as directional rather than a definitive farm census. AWS research coverage of the sector findings aboutamazon.com.au

The finding changes the interpretation of the sector. New Zealand has highly visible AI companies and several operating deployments, but these do not mean that most farms are using AI. Basic use—general-purpose assistants, digital records, cloud software or isolated decision tools—should not be conflated with advanced, integrated or autonomous operation.

AWS’s broader results reinforce the point. Only 15% of New Zealand businesses reported having a formal AI strategy, while strategy-led organisations were more likely to report advanced use and confidence in measuring returns. The implication for agriculture is that adoption will depend less on access to another tool and more on identifying a funded business problem, integrating the relevant data and assigning responsibility for results. aboutamazon.com.au

LIFT has moved from announcement to on-farm delivery

The Livestock Innovation Farming Transformation programme, or LIFT, has progressed materially since its June launch. Pāmu and its partners announced nine participating farms on 14 October, spanning Northland, the Bay of Plenty, the King Country, Whanganui, Hawke’s Bay, Marlborough, Canterbury and Otago. The group includes Māori-owned farms, high-country stations, family businesses, corporate operations and a Pāmu property. Virtual-fencing deployment is under way, and independent consultants are establishing production, financial and environmental baselines. Pāmu programme update pamunewzealand.com

This is a meaningful change in evidence quality. LIFT is designed to assess not only whether collars work, but whether farmers can change grazing systems, labour allocation, water management, environmental performance and farm profitability. Its use of FARMAX and Overseer models, along with shared data and annual evidence reports, should produce more useful information than isolated testimonials.

The first farm-level accounts are positive but preliminary. At Lochindorb Run in South Otago, a farmer reported less pasture damage, easier back-fencing and more responsive calving management after using Halter collars for approximately two months. Those observations are useful for identifying mechanisms of value, but they are not outcome measurements and cannot yet establish a return on investment. Farmers Weekly account farmersweekly.co.nz

Hectre has opened a research laboratory, but internal-fruit AI remains pre-commercial

Hectre opened an Auckland laboratory in October to develop hyperspectral imaging for assessing internal apple characteristics without cutting the fruit. The project targets ripeness, starch content, firmness and internal defects, with an algorithm proof of concept targeted for June 2027 and a packhouse prototype validated under commercial conditions by June 2028. Hectre project coverage Government project announcement revistacultivar.com

This is a progression from grant announcement to tangible research infrastructure, not evidence of a production system operating at scale. Hectre’s existing external-fruit assessment products are commercial, and the company says its technology is used in 22 countries. Those scale and accuracy claims are company-reported. The new internal-quality capability remains a research and validation project.

Halter’s scale has increased, but reliability has become part of the adoption story

Halter’s reported commercial footprint has continued to grow. In August, Farmers Weekly reported company figures of 3,500 farms and ranches worldwide and 1.5 million collars on cattle, up from the figures cited in the previous update. These are company-reported numbers rather than independently audited measures of active use or farm-level outcomes. Farmers Weekly report farmersweekly.co.nz

At the same time, public reporting of early-model collar reliability problems provides an important counterweight to the scale narrative. Halter said it was replacing affected collars and compensating or providing concessions to farms that experienced disruption. The company also described expanded reliability testing and failure analysis. Farmers Weekly report farmersweekly.co.nz

The lesson is broader than Halter. In physical agriculture, adoption depends on the reliability of the entire service: hardware, batteries, connectivity, software, alerts, support and fallback procedures. A model can perform well while the product still fails commercially if farmers cannot rely on it through winter, poor connectivity or an equipment fault.

The national AI platform remains an investment pathway, not an operating agricultural service

New Zealand’s national AI research platform remains strategically relevant to agriculture, particularly through the BioAI and Physical AI concepts focused on agriculture, forestry, outdoor environments, robotics and Māori data sovereignty. However, public evidence still does not establish a fully operating, agriculture-facing national platform with broad industry access or measurable farm deployments.

MBIE’s programme material confirms the government’s planned investment and the shortlisted concepts. A September 2026 independent tracker continued to classify funding as confirmed while treating delivery, access, utilisation and commercial outcomes as open evidence gates. MBIE platform information NZIAT evidence tracker mbie.govt.nz

Current State of AI Adoption

Livestock and pastoral systems

Livestock remains the most commercially advanced area of agricultural AI in New Zealand. The leading systems combine sensors, positioning, computer models, mobile software and farm workflows rather than presenting AI as a standalone chatbot.

Halter is the clearest example. Its platform combines GPS-enabled collars, virtual fencing, animal-behaviour monitoring, pasture intelligence, mobile control and satellite connectivity. Its Clank agent uses generative AI to investigate system alerts and assist with engineering workflows. AWS and Halter reported that the system works across more than 16,000 monitoring alerts and saved more than 215 hours of engineering time in its first months. These are vendor- and AWS-reported figures, not independent operational audits. AWS case study aws.amazon.com

The important adoption signal is not simply that Halter uses AI. It is that AI is embedded in infrastructure that farmers already use to move and monitor cattle. This reduces the need to persuade farmers to adopt an entirely new category of software, but it also raises the standard for uptime, support and accountability.

Aimer operates at a different layer. Its AIMER platform uses smartphone computer vision, satellite information and farm-specific modelling to measure pasture and support grazing decisions. The company reports use on more than 650 farms and more than 10,000 pasture measurements each week. Its government-supported expansion includes Ask AIMER, a conversational assistant intended to turn measurements into tailored recommendations. These scale figures are self-reported, and the assistant remains in development and scale-up. Aimer Farming aimer-farming.com

The emerging pattern is that AI is most useful where it is attached to a repeated decision: where livestock should graze, which areas should be protected, how pasture should be allocated, or which alerts require attention. The systems are less mature when they attempt to provide broad, farm-wide advice without reliable access to the farm’s underlying data.

Generative AI and farm administration

Generative AI remains an early-adopter technology on farms.

DairyNZ-commissioned research by Perrin Ag found that early adopters were using ChatGPT, Claude, Gemini and Copilot for drafting documents, summarising technical material, analysing spreadsheets, preparing standard operating procedures, translating information, creating farm-specific chatbots and exploring management scenarios. The report describes current use as limited and concentrated among farmers and rural professionals who are already comfortable with digital tools. Perrin Ag report perrinag.net.nz

The most credible near-term applications are administrative and interpretive:

• Drafting employment, health and safety or compliance documents.
• Retrieving information from trusted farm or industry sources.
• Translating instructions and improving communication across teams.
• Summarising records and preparing meeting material.
• Supporting scenario analysis without replacing farm or professional judgement.

The main risks are familiar: hallucinated advice, weak source control, privacy breaches and overconfidence in outputs. DairyNZ’s guidance therefore positions GenAI as an assistant rather than an autonomous decision-maker. dairynz.co.nz

Horticulture, packhouses and crop monitoring

Horticulture has the broadest pipeline of computer-vision and sensing projects, but it also shows the clearest distinction between commercial deployment and development work.

Hectre’s existing products use computer vision to assess fruit size, colour and external quality. Its internal-fruit project is now supported by a dedicated Auckland laboratory, but the published milestones show that the technology is still being trained, tested and validated. The proposed value is substantial: better information could influence storage, packing, shipping and waste decisions before fruit is damaged or downgraded. The commercial question is whether the system can maintain accuracy at packhouse throughput and across cultivars, regions and seasons.

The University of Canterbury’s UC Vision and HoloCrop are developing another form of horticultural intelligence. Their systems use cameras and AI to reconstruct three-dimensional models of plants and fruit hidden behind foliage. UC reports fruit-counting accuracy of approximately 2% to 3% during testing and says data has been collected across 20 commercial farms. These are promising research results, but the team is still working toward reliability and commercialisation through HoloCrop. University of Canterbury canterbury.ac.nz

AGMARDT has also funded HoloCrop to develop a vehicle-agnostic three-dimensional crop-scanning system for yield forecasting, harvest planning and supply-chain efficiency. neXtgen Agri has received funding for an AI-enabled ultrasound and computer-vision system to measure intramuscular fat in live sheep. Both projects are development work rather than evidence of scaled deployment. AGMARDT grants agmardt.org.nz

Precision agriculture and farm data infrastructure

Precision agriculture is an important foundation for AI, but not every digital system is itself AI-enabled. GPS guidance, proof-of-placement records, satellite imagery, sensors, automated application rates and farm-management software create valuable data without necessarily using machine learning.

The next step is integration. Farmers often have information distributed across accounting systems, production records, environmental tools, spreadsheets, machinery platforms and adviser workflows. AWS and New Zealand agritech leaders have identified interpretation and decision support—not simply more data—as the central opportunity. iStart discussion of AWS Cloud and AI Day istart.co.nz

This favours platforms that can operate offline, connect to existing equipment and return a clear next action. It also favours vendors with long-standing relationships in the sector. A new model with no integration or service pathway is unlikely to outperform a narrower tool that fits the existing farm system.

Physical AI and robotics

Physical AI—computer vision, autonomous machinery, robotics and systems that sense and act in the physical environment—has significant potential in agriculture but remains less widely deployed than software-based AI.

The sector’s use cases include:

• Livestock monitoring and virtual fencing.
• Dairy-yard control and animal-flow management.
• Orchard scanning and crop forecasting.
• Automated fruit inspection and grading.
• Targeted spraying and weed control.
• Autonomous or semi-autonomous farm vehicles.
• Biosecurity surveillance.

The latest AWS-reported research indicates that awareness and readiness remain low. Coverage of the research reported that only around 30% of agricultural businesses had heard of physical AI and that approximately 7% of those aware were using or piloting it. Because the figure comes from a vendor-commissioned survey and a broader primary-sector grouping, it should be read as an indicator of limited diffusion rather than a definitive adoption rate. Report of the AWS findings b2bnews.co.nz

Connectivity is a major differentiator. Halter’s direct-to-satellite service is intended to extend virtual fencing to remote beef farms that previously lacked adequate cellular infrastructure. The company estimates that the service could expand access among New Zealand beef farms by at least 20%; this is internal modelling, not an independently measured adoption outcome. Halter announcement halterhq.com

Governance, Policy and Regulation

New Zealand remains principles-based and technology-neutral

New Zealand’s national AI approach continues to emphasise adoption through existing law and responsible-business guidance rather than a standalone AI Act. The Government’s strategy identifies agriculture as an area where AI can enhance productivity and encourages businesses to adopt AI with confidence while managing privacy, safety, accountability and transparency risks. MBIE AI strategy mbie.govt.nz

For agricultural businesses, obligations remain distributed across:

• The Privacy Act and information-security duties.
• Consumer protection and claims about performance.
• Animal-welfare law and codes of welfare.
• Health and safety requirements.
• Environmental and agrichemical controls.
• Aviation and drone rules.
• Employment and workplace-monitoring obligations.
• Intellectual property and commercial confidentiality.

The approach provides flexibility, but it places substantial responsibility on vendors, farmers and industry bodies to define practical safeguards.

Virtual fencing remains the clearest governance test

Virtual fencing illustrates the difference between general legal coverage and technology-specific governance.

MPI has stated that wearable collars are managed under the collars and tethers provisions of the Animal Welfare Act framework, with existing welfare codes requiring animals to be handled in ways that minimise pain, injury and distress. However, veterinarians and animal-welfare organisations have argued that specific standards are needed for virtual fencing, including limits on aversive cues, monitoring of individual animals, emergency procedures, communications failures and record-keeping. Coverage of the regulatory position VAWA code of ethical conduct farmersweekly.co.nz

The issue is becoming more important as virtual fencing moves into hill-country systems and national demonstration programmes. A system that influences animal movement must have clear rules for:

• Identifying animals that do not learn or repeatedly breach boundaries.
• Preventing excessive cues.
• Maintaining access to water, shade, shelter and safe terrain.
• Responding to battery, GPS, software or communications failure.
• Allowing rapid human intervention.
• Retaining welfare records for audit or investigation.

LIFT’s commitment to monitor animal welfare and report practical trade-offs is therefore significant. It may help move debate from competing claims toward evidence about how the systems perform under New Zealand conditions.

Māori data sovereignty and participation are becoming implementation issues

The national AI platform concepts include Māori participation and Māori data sovereignty. LIFT also includes Māori-owned farms and states that farmers retain control of their data and decisions while agreed information is shared for evaluation. This is a more practical model than treating data governance as a statement of principle alone.

For agricultural AI, the relevant data may include land capability, whakapapa-linked assets, production systems, genetics, environmental information, business performance and taonga species. Governance questions include who can access the data, who can train models on it, who benefits from derived insights and whether data can be transferred across commercial or public platforms.

Evidence of mature, sector-wide arrangements remains limited. This is an area where policy, tikanga and commercial practice are still developing.

National research infrastructure is promising but not yet proven

The proposed national AI research platform could provide shared compute, datasets, test sites, safety methods and access for smaller firms. The shortlisted concepts include physical AI for outdoor environments and BioAI for agriculture, aquaculture and forestry. However, the current public evidence establishes funding and programme intent more clearly than operating capability.

The relevant test will be whether agricultural researchers and companies can access usable infrastructure, whether Māori governance is embedded in decisions about data and whether research moves into commercial systems that farmers can actually use.

Case Studies

LIFT: virtual fencing as a farm-system experiment

Technology: GPS-enabled livestock collars, virtual boundaries, animal monitoring, pasture management and data-supported grazing decisions.

Operating status: Nine commercial farms have been selected and onboarded. Collaring and training are under way, with baseline production, financial and environmental models being established. Pāmu LIFT update

Evidence status: Early farmer accounts are positive but anecdotal. The programme’s value will depend on future evidence reports, cross-farm comparability and transparent reporting of failures and trade-offs.

Strategic significance: LIFT is one of the strongest attempts in New Zealand to generate independent, system-level evidence for agricultural AI and physical automation.

Halter: operating physical AI with a reliability burden

Technology: Smart collars, virtual fencing, animal behaviour monitoring, pasture intelligence, satellite connectivity and agentic AI for platform operations.

Operating status: Commercially deployed in New Zealand and overseas. Halter has reported 3,500 farms and ranches and 1.5 million collars globally. Clank is operating within the company’s engineering workflows. These figures are company- or AWS-reported. Scale report AWS case study

Evidence status: The platform is clearly operating at commercial scale, but independent evidence of farm-level financial and environmental returns remains limited. Public reporting of early collar reliability issues shows that operational resilience is part of the adoption case.

Strategic significance: Halter demonstrates how AI becomes agricultural infrastructure when connected to an existing farm workflow. It also shows why physical-AI companies must be judged on the full service, not on model capability alone.

Aimer Farming: from measurement to recommendation

Technology: Smartphone computer vision, pasture measurement, satellite data, farm modelling and a planned conversational assistant.

Operating status: Aimer reports use on more than 650 farms, with more than 10,000 pasture measurements each week. Ask AIMER remains part of the government-supported expansion programme. Aimer Farming

Evidence status: The platform’s operating scale is self-reported. Independent evidence of average farm-level returns is not yet publicly established.

Strategic significance: Aimer illustrates the importance of low-friction, offline-capable tools that fit existing pasture-management routines.

Hectre: commercial external inspection, experimental internal analysis

Technology: Computer vision for fruit sizing and external quality, with hyperspectral imaging under development for internal fruit characteristics.

Operating status: Hectre’s external-fruit products are commercial. The internal-quality system is in laboratory and proof-of-concept development, with commercial validation targeted for 2028. Hectre says its systems are used in 22 countries and process billions of pieces of fruit annually; these are company-reported figures. Hectre laboratory report

Evidence status: Existing external-inspection products have a stronger deployment case than the new internal-analysis project. The latter remains subject to accuracy, throughput and integration tests.

Strategic significance: Hectre shows how New Zealand agritech can use the domestic packhouse environment as a research and commercialisation base for export products.

DairyNZ and Perrin Ag: cautious generative AI adoption

Technology: General-purpose language models, custom farm chatbots and AI-assisted document, knowledge and data workflows.

Operating status: Early-adopter use rather than broad sector deployment.

Evidence status: The research provides useful qualitative evidence of current practice, but not a statistically representative estimate of dairy-farm adoption or returns.

Strategic significance: The most immediate value of GenAI may be reducing administrative load and improving access to trusted information. The technology is more likely to augment farm teams than replace agricultural judgement in the near term.

Trends

1. Commercial success and farm-gate adoption are diverging

New Zealand’s agritech companies are attracting capital, exporting products and operating sophisticated AI systems. Yet the latest adoption evidence suggests that most agricultural businesses are not using AI at an advanced level.

This is not necessarily a contradiction. A small number of well-funded companies can be globally advanced while their customer base adopts only specific features. The sector’s challenge is to translate company-level capability into repeatable, affordable and supportable farm-level use.

2. The next adoption wave will be judged by evidence, not demonstrations

LIFT is important because it creates a structure for measuring production, financial, environmental and practical outcomes across different farm types. Hectre’s staged milestones similarly distinguish algorithm development from commercial validation.

This is the direction the sector needs: explicit baselines, pre-defined success measures, independent analysis and public reporting of negative results as well as positive ones.

3. Reliability is becoming a product feature

Hardware failures, battery performance, network availability, data loss and weak support can destroy confidence in otherwise capable technology. The Halter collar issue is a visible example, but the principle applies to cameras, drones, sensors, automated gates and robotic machinery.

Agricultural AI must work through weather, dust, terrain, seasonal change and inconsistent connectivity. Reliability engineering may therefore create more competitive advantage than another marginal improvement in model accuracy.

4. Offline-first and edge-capable systems have a structural advantage

AIMER’s smartphone-based approach, rural satellite connectivity and research into low-power physical AI all reflect the same operating reality: many New Zealand farms cannot assume continuous high-bandwidth connectivity.

Tools that can capture data offline, synchronise later, provide clear local alerts and continue operating safely during outages are more likely to gain trust than systems designed around permanent cloud access.

5. Data integration is more valuable than isolated prediction

The strongest commercial opportunities are moving toward systems that connect pasture, livestock, weather, soil, machinery, environmental and financial data. Farmers do not generally need another dashboard; they need a reliable interpretation of information already being generated.

This creates an opportunity for farm-management platforms, advisers and rural service providers that can combine data without taking control away from producers.

6. Physical AI is advancing faster in capability than in adoption

Virtual fencing, computer vision and autonomous machinery are becoming technically credible. Their deployment is still constrained by cost, skills, connectivity, regulation, animal welfare and the need to change farm routines.

The gap between a system working in a controlled demonstration and a system working across many farms remains substantial. New Zealand’s varied landscapes make it a valuable test environment, but they also make generalisation difficult.

7. Export orientation remains central

The domestic market is too small to support many agritech businesses without international growth. Halter, Hectre and other companies are therefore using New Zealand farms, orchards and packhouses as testing grounds for systems designed for wider markets.

This creates a strategic benefit for New Zealand but also a risk: products may be optimised for export growth rather than for affordability and interoperability across smaller domestic operations.

Outlook

Over the next two years, the most credible pathways for AI adoption are likely to be:

1. Evidence-led expansion of virtual fencing
   LIFT should produce the sector’s clearest evidence yet on where virtual fencing changes farm performance, where it does not, and what it costs to operate.

2. More AI within established farm platforms
   Existing systems with customer relationships and data access are likely to add conversational interfaces, predictive alerts and workflow automation faster than entirely new autonomous farm platforms will scale.

3. Commercial tests for horticultural sensing
   Hectre’s 2027 proof-of-concept and 2028 commercial-validation milestones will provide a useful test of whether hyperspectral internal-fruit assessment can move from laboratory research to packhouse infrastructure.

4. Continued growth in administrative GenAI
   Farm documentation, technical search, translation, training and scenario analysis should grow faster than unsupervised systems making animal, chemical or environmental decisions.

5. More attention to operational resilience
   Vendors will need to demonstrate service levels, fallback modes, data recovery, human override and support processes alongside model accuracy.

6. A sharper divide between prepared and unprepared businesses
   Farms and agribusinesses with clean data, reliable connectivity, capable staff and a defined business problem will be better positioned to capture value than those buying disconnected tools without a change plan.

The major uncertainty is distribution. AI could lift productivity in data-rich, well-capitalised operations while leaving smaller farms, remote producers and less-connected regions behind. Broad adoption will require practical extension, finance, interoperability and independent evaluation—not only more funding for prototypes.

Overall Assessment

AI in New Zealand agriculture is commercially credible, operational in selected systems and still far from sector-wide.

The strongest evidence remains in livestock and pastoral applications. Halter is operating a connected physical-AI platform at meaningful scale, while Aimer is turning pasture measurement into increasingly actionable decision support. LIFT has now created a structured national test of virtual fencing in sheep and beef systems, although robust outcome data is still ahead.

Horticulture has a strong pipeline, with Hectre’s laboratory opening and UC Vision’s three-dimensional crop models showing that New Zealand’s capability extends beyond livestock. However, much of this activity remains in research, proof-of-concept or commercial-validation stages.

Generative AI is spreading more cautiously. Its immediate agricultural value lies in administration, knowledge retrieval, communication and scenario analysis. There is little evidence that general-purpose AI has yet become a mainstream farm-management system.

The latest adoption data is a useful corrective to industry hype. New Zealand can produce globally ambitious agricultural AI while most farms remain basic users or non-users. The strategic question is therefore no longer whether New Zealand can build impressive systems. It is whether those systems can be made reliable, affordable, interoperable and useful across ordinary farms in ordinary seasons.

The technologies most likely to endure will combine agricultural knowledge with robust infrastructure, transparent data practices, human oversight and independently credible evidence of value.