Navigating AI deployment in precision livestock farming: current trends and future prospects

Abstract

The selection of an AI deployment model is a critical strategic decision for livestock operations, as no single solution fits all scenarios. Cloud-edge collaborative architecture is emerging as the most effective paradigm, balancing on-farm responsiveness with powerful cloud analytics. Widespread AI adoption relies on overcoming key real-world barriers, including rural connectivity, implementation costs, and the on-farm technical skills gap. Future PLF advancements will depend on integrating multi-modal data to create more holistic and prescriptive animal health and welfare management systems. Precision livestock farming (PLF) is undergoing a profound transformation, with its core driver shifting from traditional data collection to intelligent decision-support systems powered by artificial intelligence (AI). While early-stage PLF relied on simple sensors for discrete tasks like estrus detection, rapid advancements in the Internet of Things (IoT), sensor technology, and computing power now enable modern systems to gather vast, multidimensional data covering animal behavior, physiology, and their micro-environment (Alexy & Haidegger, 2022; Kaur et al., 2023). This evolution is driven by multiple pressures facing the global livestock industry: rising labor costs and shortages compel farms to seek automation for efficiency, while increasing consumer and regulatory demands for product quality, animal welfare, and sustainability necessitate more refined management methods. Reflecting this momentum, the global PLF market is projected to expand at a compound annual growth rate of over 10% through the next decade, signaling strong and sustained industry adoption (Sojitra & Dudhagara, 2023). Consequently, AI’s role has evolved from a frontier concept to an indispensable engine for industry advancement. The proliferation of data has catalyzed a surge in academic research focused on developing sophisticated AI algorithms to enhance livestock production, health, and welfare (He et al., 2025). These studies have demonstrated significant potential, with models capable of predicting metabolic diseases (Giannuzzi et al., 2023), detecting specific behaviors with superhuman accuracy (Kang et al., 2020), and optimizing feeding strategies (King et al., 2024). However, the majority of this research has concentrated on algorithmic innovation and validation in controlled environments. A critical gap persists between the development of high-performing algorithms and their practical, scalable, and robust implementation on commercial farms (Berckmans, 2017). The crucial questions of how these AI systems are deployed, the architectural trade-offs involved, and the real-world challenges encountered often remain underexplored. This disconnect hinders the translation of technological potential into tangible on-farm value. This review offers an insightful overview and future perspective on the primary AI deployment pathways in PLF, with a practical, application-driven approach. We will systematically dissect the mainstream architectures, including offline analysis, on-premises servers, edge computing, cloud platforms, and emerging cloud-edge collaborative frameworks. By examining the inherent advantages, limitations, and practical trade-offs of each pathway through recent case studies, this review will illuminate the critical challenges hindering widespread adoption, such as latency, connectivity, and data privacy. Ultimately, this article will offer a forward-looking perspective on future ­developments, providing valuable guidance for researchers, technology developers, and industry practitioners working to build the next generation of effective and accessible AI solutions for modern livestock farming. The deployment of AI in PLF is not a monolithic practice but exists along a diverse spectrum. This spectrum ranges from fully farm-controlled, capital-intensive on-premises systems to highly flexible, service-dependent cloud solutions, with various hybrid models in between. The selection of a deployment model is therefore not merely a technical decision but a strategic one, reflecting a farm’s operational scale, capital resources, technical capabilities, and philosophy on data as a core asset. This decision-making process is an intricate exercise in trade-offs. For instance, a small family farm with limited capital and no specialized IT staff is unlikely to build and maintain an expensive on-premises server, which demands significant upfront investment and continuous professional oversight. For such operations, low-cost, user-friendly mobile applications or pay-as-you-go cloud services represent a more realistic and accessible entry point (Shwetabhand & Ambhaikar, 2024). Conversely, a large, vertically integrated agricultural corporation may view its farm data as a key competitive advantage. Driven by concerns over data security, privacy, and ownership, and to comply with stringent internal governance or regional regulations, such an enterprise would likely invest in a private on-premises or hybrid system to ensure sensitive data never leaves the farm’s physical or virtual perimeter (Moreira et al., 2024). Geographical location and infrastructural conditions are equally decisive factors. For farms in remote areas with unstable or limited internet connectivity, a purely cloud-dependent solution is unfeasible. In these scenarios, edge computing or a cloud-edge collaborative architecture, which can perform critical data processing locally, becomes a necessity for ensuring system reliability (Batistatos et al., 2025). Consequently, a nuanced understanding of the logic and trade-offs inherent to each deployment pathway is essential. The critical consideration shifts from identifying a universally “best” technology to selecting the most suitable architecture for a specific operational context. This section provides a systematic analysis of the five mainstream deployment pathways along this spectrum, which are visually summarized in Figure 1. Comparative analysis of mainstream AI deployment pathways in precision livestock farming. The table evaluates the various deployment models across six key dimensions: platform, cost, latency, security, scalability, and dependency. Example of offline AI analyses performed on precollected datasets for production forecasting and health monitoring. (a) Ji et al. (2022) show the prediction of future milk yield from historical records. (b) Kang et al. (2020) and (c) Jiang et al. (2022) demonstrate different computer vision approaches for post-hoc lameness detection, analyzing back curvature and hoof supporting phase from video data. Examples of on-premises AI deployment for real-time monitoring. (a) Jung et al. (2021) illustrate a system where audio data from microphones is processed on a local PC for cattle vocalization analysis. (b) Huang et al. (2023) show a vision-based system where camera data is transmitted to an in-house server for real-time cow tail tracking. Example of studies using edge deployment for real-time, on-device animal monitoring and health diagnostics. (a) Zhou et al. (2024) show the workflow for swine behavior analysis using a Jetson Nano; (b) Xiao et al. (2024) illustrate a system for cow identification on a Jetson Xavier NX; (c) Aravamuthan et al. (2024) detail a portable device for digital dermatitis detection; and (d) Kingsley et al. (2025) presents a mobile application for goat disease detection. Examples of cloud-based deployment architecture for scalable livestock monitoring. (a) Unold et al. (2020) illustrate a general cloud system, while (b) Dineva and Atanasova (2021) and (c) Bhaskaran et al. (2024) showcase specific scalable architectures built on Amazon Web Services (AWS) for smart livestock management and real-time health alerts. Examples of cloud-edge collaborative deployment architecture. (a) Srinivasagan et al. (2025) illustrate a workflow where a model is trained in the cloud and deployed on a low-power edge device for real-time inference. (b) Shen et al. (2021) show a system where the edge device performs local data processing and classification, sending only the results to the cloud for long-term aggregation. Offline deployment represents a foundational and widely adopted paradigm for applying AI in PLF, characterized by its “collect-first, analyze-later” approach (Figure 2). In this pathway, farms systematically accumulate data over extended periods, forming comprehensive historical datasets that are subsequently used to train and validate machine learning models in a nonreal-time environment. This decoupling of model development from daily farm operations allows for deep, retrospective analysis aimed at informing long-term strategic decisions rather than immediate interventions. This deployment model is prevalent in academic research and has been successfully applied to address key challenges using various data types. For tabular and sensor data, offline models have demonstrated significant predictive power. For example, Perneel et al. (2024) successfully explained up to 47% of the variance (R2) in a cow’s lifetime production potential by analyzing historical genetic and environmental records using stacking ensemble models. Similarly, a random forest model developed using 20 years of test-day records was able to forecast early-lactation milk yield with a root mean square error between 6.08 and 6.24 kg (Salamone et al., 2022). In health applications, high accuracy has been achieved in predicting blood metabolites from milk infrared spectra (Giannuzzi et al., 2023), while other models have effectively predicted insemination outcomes (Shahinfar et al., 2014) and forecasted future milk yield (Ji et al., 2022). Vision-based analysis is another prominent domain for offline deployment, where extensive video or image data is processed post-hoc (Oliveira et al., 2021). In lameness detection, for instance, Jiang et al. (2022) developed sophisticated deep learning pipelines that combine custom object detection with network models to classify lameness from back curvature data with 96.61% accuracy. Another approach analyzed the hoof supporting phase using a deep learning network, resulting in 96% classification accuracy (Kang et al., 2020). This method has also proven effective for monitoring feeding behavior, as a study by Bresolin et al. (2023) trained a YOLOv3 deep learning model on annotated historical images to achieve 96.0% accuracy in individual heifer identification, which in turn enabled the precise calculation of feeding time (R2 = 0.99). A primary advantage of offline deployment is its minimal requirement for on-farm real-time infrastructure, which lowers the barrier for adoption. It allows for the use of large-scale, longitudinal datasets and computationally intensive algorithms to build robust models that support strategic planning. However, the principal limitation of this pathway is its inherent lack of real-time actionability. Models cannot provide immediate alerts for acute health events, and insights are generated retrospectively, meaning the optimal window for intervention may have already passed. Consequently, models trained exclusively on historical data may become less accurate as farm conditions evolve, positioning this approach as a reactive, rather than a real-time management deployment a significant from offline analysis real-time farm management (Figure This pathway local computing such as or the farm’s from or microphones is transmitted over a local network to these in-house for immediate The core advantage of this model is its to AI algorithms in real-time or immediate alerts and management insights on an internet for the primary This monitoring and interventions. The application of on-premises systems is valuable for monitoring where immediate is A of this architecture can in where a system used microphones in a to cattle The to a on-premises PC that a network in real-time, up to accuracy in cattle from and providing alerts for like estrus or et al., 2021). In vision-based behavior and health the of on-premises deployment on the of the AI models. studies have developed powerful algorithms suitable for this For instance, high for and for have been achieved in goat behaviors using a system that on the of a local server et al., 2020). systems for have of over et al., and an model for cow tail achieved in a the high of this model its for on-premises deployment, where provide real-time for or health et al., 2023). The primary of on-premises deployment is its for immediate to This model also offers data and security, as all is processed and operational of unstable rural internet However, this approach with significant The capital for server can and these systems technical also presents a as the system often in to the more and by cloud-based deployment represents the most and pathway for AI in PLF, where intelligence is from the (Figure In this AI algorithms on such as systems integrated into smart or mobile This allows for on-device processing of data at the the for continuous data to a local server or the cloud and The of edge deployment on the development of highly AI models that can perform tasks with minimal advancements have this for real-time monitoring. In swine behavior analysis, for example, Zhou et al. (2024) successfully deployed a model for detection and a model for a Jetson with the model achieved an processing of on the edge while the video analysis model processed in Similarly, real-time individual cow identification at 20 has been achieved by a and a custom model on a Jetson Xavier et al., 2024). deployment is also effective for real-time health diagnostics. A portable system developed by Aravamuthan et al. (2024) for detecting digital dermatitis a model on a Jetson Xavier a rate of with a of its The of edge AI to and mobile A custom for cattle for instance, at image on a et al., 2022). have successfully deployed applications on by models into like Examples a model for digital dermatitis detection on and as as 20 on an et al., and a system using for goat disease detection et al., 2025). The primary advantage of edge deployment is its latency, which immediate alerts and real-time By processing data locally, this approach network to operational It also data and as sensitive not to the and high operational reliability in areas with or no internet However, this pathway also presents significant The of edge are a primary as AI models and often a between and predictive accuracy. This process specialized can an is a large, significant and to deployment represents a paradigm where data from on-farm and other are transmitted the internet to remote server for and analysis (Figure on-premises solutions that on local farm servers, this model the and by cloud like Amazon Web Services and By shifting the core tasks of data processing and AI model from the farm to the this approach scalability, robust data and the application of algorithms to deep insights for farm The evolution of cloud technology has as a solution for the and datasets inherent in comprehensive livestock health monitoring systems built on & 2021). In these sensors key such as and the data to the A of including for device management and Amazon for in while machine learning models on Amazon the data to health For instance, an is such as a rate a of the system can an to a mobile device et al., 2024). Another application by et al. (2024) the cloud for the detection of in through data This successfully data images for sensor data, and The workflow is in the with the image processing a of for accuracy for image and for individual animal By deep from these the predictive model achieved an as high as the power in integrating data for disease The primary advantage of cloud deployment is its scalability, farms to expand from a to of sensors upfront investment et al., 2024). It provides the power for sophisticated deep learning models and as a for data remote to and alerts et al., 2020). However, this model is on which can a significant barrier in rural costs, for real-time applications, and concerns over data on remain key challenges to its widespread adoption. Cloud-edge collaborative deployment represents a sophisticated hybrid architecture that tasks between on-farm edge and cloud (Figure This model has as a solution to the challenges of latency, network limitations, and data in a cloud model where edge merely as data in this the edge performs The implementation of this is by how tasks are between the edge and the approach data and at the edge to data In a system for monitoring cow for example, edge process data, a to classify behavior in the classification a simple is to the cloud for long-term aggregation. This method the data by an while a local classification accuracy of et al., 2021). A more pathway the cloud for intensive model and while the resulting models for real-time at the This is in studies on vocalization and For welfare analysis, a model is trained and on a cloud deployed on a low-power device in the an as as with accuracy et al., 2025). Similarly, for a is trained in the and its intelligence is to a through deployed on an Jetson edge platform, this model time by to et al., 2025). The most of real-time on In a cattle identification an model on an Jetson edge the prediction is high the classification is However, the is a the is to a more powerful model in the cloud for analysis. This workflow tasks are at the while decisions from the an system accuracy of & 2025). While this hybrid model offers significant advantages, including latency, network and data privacy, its practical implementation presents The of a system is a primary robust for device monitoring and effective is critical for data and by The of model is also This process technical pathways to models for edge often model to or and such as model and to a These a high of technical and represent to the deployment of cloud-edge systems in real-world farm et al., 2025). The of AI from controlled research to the of commercial farms is with significant practical These represent the gap between the potential of the deployment models and their real-world and systematically to widespread adoption. the most the practical deployment of AI is by the infrastructural and environmental of the The of and cloud-edge collaborative architectures is on internet that in rural agricultural This digital deployment and tasks like model et al., 2023). the physical conditions of characterized by and robust and for on-premises and edge to and the physical infrastructure, a significant technical skills gap presents another operational The deployment and of AI sophisticated cloud-edge specialized in IT and machine learning This in to the general on most a on continuous The for management to tasks like model and is therefore critical for this gap. technical and infrastructural challenges are and a on investment remain a for While the technical of AI models can is a lack of comprehensive studies these into tangible such as in or The capital for like Jetson with operational costs for cloud a case that is not fully with data et al., 2024). these practical are core concerns of data privacy, security, and operational data a valuable and sensitive commercial asset. Consequently, cloud-based can data and the potential for by While on-premises and edge offer data the and of across the data are crucial for the for adoption. data the deployment of AI in PLF that are to the architecture. For example, cloud-based models that data from farms algorithmic datasets are not to or for farm types. Another is the of & 2022). to welfare, AI systems used to to their for the high upfront with sophisticated on-premises or edge systems a technological between and family These consideration in the and deployment of AI solution in the livestock Ultimately, the of deployed system on and A sophisticated AI system is of its are not or by farm The of and the of its insights are demonstrated by and a that time by can in overcoming this these challenges is crucial for AI in PLF from research to on the of AI deployment, the future of PLF is for an evolution from applications an intelligent This will by key that address challenges and The most significant will a the of single sensor monitoring the of diverse data from like that of et al. which and data, the next generation of AI will a holistic digital for each By these such as and data, these systems can simple detection to provide deep, insights into an health, welfare, and the role of AI will from predictive alerts to prescriptive decision This an evolution from systems that only potential to that than a for a of future systems will provide management such as a hoof on a in an This AI from a monitoring into an intelligent for and farm The of this technology will also driven by primary the of powerful and edge such as the used by Srinivasagan et al. will sophisticated on-farm AI more the of like will This approach allows AI models to trained across multiple farms the of sensitive data, the core data and innovation et al., 2023). However, technology cannot the adoption gap. A adoption is crucial for of all to For small and this may with solutions like offline analysis of records or a single edge device to address a critical point and validate agricultural can a and different deployment architectures in specific to a capital-intensive et al., 2017). This approach that technology adoption is driven by proven but its also on a will for agricultural services data into and AI is to farm the will from systems to scalable The model of systems will to and that This is already catalyzed by the proliferation of powerful like et al., and et al., which provide a will to from various data and a more competitive and market that can from small to large, This article has the of AI in PLF, its from a concept into a core engine the The systematic analysis of the five primary deployment pathways a and while each model offers advantages, the cloud-edge collaborative architecture has as the most and practical to real-time responsiveness at the edge with the power of the cloud to the demands of modern livestock However, the to widespread adoption is on overcoming significant to rural infrastructure, technical a and critical concerns over data by researchers, developers, and industry These are not but that will the future of the potential for AI in this domain is The of data prescriptive decision and will a generation of intelligent farming. This future not only in production but also profound in animal welfare and environmental By to the gap between algorithmic potential and on-farm AI is to the future of livestock a system that is more and is a in the of and at the of and of and in computer and technology from of and the of in and research on the application of computer vision and machine learning for animal behavior and welfare, as as the development of cloud-edge collaborative systems. is a in the of and at the of and of and in and intelligent system from and in and research on computer deep precision livestock artificial intelligence for animal behavior, and is an of in the of of the of at research on strategies to enhance efficiency, environmental and animal welfare, to the long-term sustainability of farming systems. in and data to research that with real-world research strategies for and approach to understanding of and in is an in precision livestock farming and animal behavior and welfare at the of and of in and from and a in from In on a research at the and at the of an agricultural and to animal is to applying strategies and solutions to address global challenges in animal production and with a on the to animal welfare and This was by the the of The in this are of the and not the or of the the or the analysis, & & and & of The no or of The would like to the support from the at the of

Topics & Concepts