Impact of Fourth Industrial Revolution (4IR) Automation on Agricultural Employment in South Africa

Abstract

The Fourth Industrial Revolution (4IR) has introduced modern, high technologies that are automated, such as precision farming, to enhance agricultural production. However, this comes at the cost of human labor being replaced by machines that are deemed efficient. This study investigated the impact of 4IR automation on agricultural employment in South Africa, spanning from 1990 to 2024. To analyze this, the study employed the Johansen test for cointegration and the vector error correction model to test for long-run and short-run dynamics. Stationarity was achieved, and the Johansen test confirmed cointegration. The vector error correction model results revealed that both long-run and short-run relationships between 4IR automation and agricultural employment exist, indicating that human labor is particularly at risk of being replaced by automation, such as advanced agricultural machinery. The results imply that, although automation improved agricultural productivity, it caused an increase in agricultural unemployment within South Africa. Therefore, to balance the advancement of technology and agricultural employment, the study recommends skills improvement and government intervention for enhancing human labor within the agricultural sector.

1. Introduction

Automation technologies refer to systems that are used to replace human labor with machines powered to efficiently perform certain tasks within an economic environment (Sostero, 2020). In the past two decades, developments of tools such as robots used in industries and Artificial Intelligence (AI) have been on the rise, allowing machines to carry out complex work that was previously performed by humans (Arntz et al., 2020; Filippi et al., 2023).

Automation that is driven by the Fourth Industrial Revolution (4IR), including robotics, Internet-of-Things (IoT), Artificial Intelligence (AI), machine learning and digital platforms, has the capabilities to replace old tasks and create better opportunities. However, the rapid expansion of automation presents both opportunities and associated risks. Automation within various industries, including agriculture, boosts efficiency using high-tech innovations (Anzolin, 2021). However, this comes at the expense of jobs while broadening income inequality (Guo, 2023). Developing countries such as South Africa face persistent unemployment and inequality despite advanced agricultural, industrial and digital infrastructure. Statistics South Africa reported a high unemployment rate and structural labor market challenges in the fourth quarter of 2024, stretching to 2025 (Statistics South Africa, 2024), showcasing the urgency of acting against unemployment.

Although automation can increase productivity, it has also reduced routine labor, raising concerns about employment outcomes and sustainability. Within the agricultural sector, where major production relies on technology, challenges of unemployment persist, particularly in both developed and underdeveloped countries. Other challenges posed by automation within the agricultural sector have been noted by several authors. For example, Sreekantha (2016) states that small agricultural land holdings are unsuitable for large agricultural machines, and the lack of financial support to purchase new advanced machinery is viewed as a setback in the world of the 4IR. In South Africa, it was found that 4IR technology use has not been equally beneficial to emerging farmers due to various constraints, such as a lack of funding (Mtshali, 2024). Thus, understanding the effects of the 4IR on agricultural employment is important for enhancing skills.

According to OECD (2020), automation restructures the labor demand by limiting the need for low-skilled laborers while high-skilled laborers become highly demanded. Frey and Osborne (2017) outlined that the advances in AI and machine learning have broadened the different activities that can be performed by automation. For instance, recent studies demonstrate that new agricultural technologies, such as those that can detect plant diseases and pests, robotic milking machines and those that improve value addition, are considered efficient (Tian et al., 2020; Upadhyay et al., 2025).

Trends in Agricultural Development

Agricultural development across the globe has swiftly moved from traditional labor-intensive practices to automation. For instance, there has been a shift from traditional ways of production to the use of new technologies that involves digitalization and automation. Previously, manual labor tools such as hand-hoes or primitive farming that involved the use of animal power were in demand and increasingly in use, particularly from 1750 to 1850 (Bazargani & Deemyad, 2024). These resources allowed for the use of manual operation, increasing the time to complete the task of production; however, they offered employment.

Between 1900 and the 1960s, agriculture saw a change in how production was performed. The introduction of machinery such as tractors during this phase was done to increase production, given the growing population and the demand for food (Bazargani & Deemyad, 2024). However, this was the beginning of the reduction in manual labor. Additionally, these agricultural machineries had the potential to provide numerous successful cropping cycles per year, thus enhancing the production capacity of farms (Peng et al., 2022).

From the 1960s, agricultural mechanization began to be improved through automation with advanced computing and electronics, which led to the improvement of productivity. The digital transition of agriculture has played a pivotal role in the transformation of agriculture, particularly through automation. Currently, new machines are integrated in agricultural production, aligning with the global use of the 4IR to improve life (Mtshali, 2024). However, this comes at the cost of replacing human labor with machines, further affecting the livelihoods of farm laborers and their households.

2. Literature Review

The South African agricultural sector dualistically comprises capital-intensive commercial farms and labor-intensive small-scale/subsistence farming; however, the latter has the largest agricultural land, covering over 96 million hectares, further contributing to the GDP by 3% (StatsSA, 2025). Moreover, the agricultural sector in South Africa employs over 900,000 people (DALRRD, 2025). This makes agriculture an important employer for many households in South Africa. Nevertheless, the sector is not exempted from 4IR automation.

For instance, Mtshali (2026) outlines that the 4IR offers opportunities to transform subsistence agriculture through the introduction of technologies such as IoT sensors, mobile advisory services, and irrigation systems that are powered by solar and digital trading platforms.

Given this, South Africa has adopted precision agriculture within the agricultural sector to integrate agronomic principles and technology to improve crop production (Wadiwala & Schoeman, 2022). However, the opportunities posed by the 4IR are costly, posing a challenge to resource-poor farmers.

Studies across the globe show that automation can be both positive and negative for employment. Habiyaremye et al. (2024) outlined that automation promotes the need for skilled laborers who can operate, maintain and interpret data from advanced technological equipment. In countries such as South Africa, the literature indicates that the use of modernized AI machines and digital farming is more predominant in large-scale farming (Habiyaremye et al., 2024). Emerging farmers who face financial constraints have limited benefits due to a lack of funds, poor extension services and a lack of training, which reduces farmers’ adoption of technology.

Drones used in precision agriculture have also been reported to improve the management of pests and diseases through recognizing damaged crops and infestation of pests and monitoring the farm (Tyagi & Pandey, 2024). Numerous studies recognized the importance of drones in precision agriculture as one of the ways to cut production costs and support research-driven decision making (Takhumova et al., 2025). In addition, the inclusion of robotics in agriculture is expanding agricultural activities through automation (Morar et al., 2020). This is because farmers use robots to consistently track and monitor crops and livestock and enhance soil precision.

South African 4IR analysis emphasizes both opportunity and risk. The Department of Trade, Industry and Competition (DTIC) promotes 4IR programs to support and strengthen industrial competitiveness (DTIC, 2025). Previously, farmers relied on visual examinations of crops to detect infections (Khakimov et al., 2022). However, this method has various limitations, such as inconsistent monitoring and not being cost-effective, and the results are likely to be incorrect (Wang et al., 2022; Minhans et al., 2025). Hatuwal et al. (2020) discovered that accurate diagnosis of diseases on plant leaves through visual observations is usually difficult, especially when the leaves are small. Therefore, the 4IR plays an important role in enhancing the identification and control of pests and diseases in agriculture.

Habiyaremye et al. (2024) further examined the use of 4IR technologies in South African agriculture. The authors outline that in South Africa, commercial farmers use 4IR technologies such as Global Positioning Systems (GPSs), Geographic Information Systems (GISs), Internet-of-Things (IoT), and drones to improve production. Moreover, the dualistic nature of South African agriculture positions large-scale commercial farming better for acquiring and adopting new technologies, as small-scale farmers are resource-constrained. In addition, AI, IoT and other technologies are employed to manage digitally integrated agricultural farms, which are mostly commercial (Bacco et al., 2019).

Although the 4IR has negative implications on employment through the reduction in labor, Habiyaremye et al. (2024) argue that other than job losses, automation frequently causes reallocation of labor and skill upgrading. Mbandlwa (2020) contends that though the 4IR might create employment opportunities for some people in South Africa, particularly high-skilled workers, it will have a negative impact on less-skilled laborers as their work can be replaced by machines.

Nkosi and Agholor (2021) conducted a review study on the 4IR and its implications on advisory services in South Africa. The paper explains the importance of agricultural extension in facilitating the adoption of 4IR technologies among emerging farmers; however, extension officers also require training and skills to effectively advise farmers, as they are their closest source of information. This makes the South African agricultural extension services important for integrating 4IR technologies in farming activities.

Studies have noted that while automation can improve productivity, it must be matched with policies that protect human labor. Some studies have noted that the 4IR presents opportunities for farmers to expand their income through the employment of smart technologies and further increase their income (Santiteerakul et al., 2020). It is against this background that this study sought to investigate the impact of Fourth Industrial Revolution (4IR) automation on agricultural employment spanning from 1990 to 2024, focusing on testing the functional relationship between agricultural employment and the usage of automation within the agricultural sector to test the long-run and short-run correlation between 4IR automation and agricultural employment in South Africa.

3. Materials and Methods

For the econometric analysis, the study used a time series, secondary data from South Africa collected from 1990 to 2024. The data was sourced from The World Bank and Our World in Data. The variables used for the analysis are listed and described in

Table 1. Data description.

	Variables	Description	Measurement	Source of the Dataset
1	AE	Agricultural Employment	% of total labor force	The World Bank
2	ATM	Automation	% of farm machinery per unit of agricultural land	Our World in Data
3	RP	Rural Population	% of total population	The World Bank
4	AL	Arable Land	% of land area	The World Bank
5	GDPPC	GDP Per Capita	Annual %	The World Bank

.

To produce an empirical study, a functional relationship needs to be established between agricultural employment and the usage of automation within the agricultural sector, as influenced by the study of Bazargani and Deemyad (2024). The current study captures the recent shift in agricultural employment with increasing usage of automation within the South African agricultural sector. The analytical model employed in the study is expressed in Equation (1):

AE_t = β₀ + β₁ATM_t + β₂RP_t + β₃AL_t + β₄GDPPC_t + µ_t

(1)

where β₀ is the intercept; β₁, β₂, β₃, and β₄ are coefficients of explanatory variables; and µ_t is the error term.

In summary, AE represent agricultural employment, ATM is the farm machinery per unit of agricultural land representing the proxy of automation, RP represent rural population, AL represent arable land, and GDPPC represents GDP per capita.

3.1. Stationarity

Testing for stationarity is important for accurate econometric forecasting to avoid biased results from non-stationarity, as non-stationarity can lead to spurious regression (Van Greunen et al., 2014). This study employed the Augmented Dickey–Fuller (ADF) test, developed by Dickey and Fuller in 1979, which tests the null hypothesis of the presence of a unit root in a time-series variable and can be used with serial correlation in the model by including lagged variables on the right-hand side, and the error terms converge closer to zero.

The ADF test for unit root estimates the following equation:

$$∆X_t={\phi }_0+{\phi }_{1}t+{\phi }_2{X}_{t-1}+{\sum }_{i=1}^n{p}_i∆X_{t-1}+{\mu }_t$$

(2)

where $∆X_t$ is the first-difference operator, n is the number of lagged differences, t is a linear trend, and µ_t is the error term.

Phillips and Perron (1988) generalized the Dickey–Fuller procedure, allowing for mild assumptions regarding the error distribution. The Phillips–Perron (PP) test addresses serial correlation by estimating the long-run variance in the error process. Although the PP test is generally more powerful than the ADF test, it is non-parametric, requiring no model or lagged parameters to be specified.

The Phillips–Perron test is represented as follows:

$$∆y_t=\gamma +{\beta }_t+\delta y_{t-1}+{\mu }_t$$

(3)

where $∆y_t$ is the first-difference representation of the series being tested, γ is a constant, β is the coefficient of t, and δ is the lag order. When there is a difference between the results of the Augmented Dickey–Fuller (ADF) and Phillips–Perron (PP) tests, the PP test is often considered more reliable. This is because the PP test accounts for serial correlation and heteroskedasticity in the error terms, making it robust under various conditions.

3.2. Johansen Test for Cointegration

Before cointegration was determined, the study outlined the criteria for selecting lag length, which included the Akaike Information Criterion (AIC), Schwarz Information Criterion (SIC), Hanna–Quinn (HQ) criterion, Final Prediction Error (FPE), and the Sequential Modified Likelihood Ratio test statistics. In this study, AIC was selected due to the small sample size of less than 60 observations. The Johansen cointegration was then employed to test for cointegration, which occurs when two or more variables share a long-run relationship, and the 1988 Johansen procedure allows for multiple economic variables. Johansen (1991) provided trace and maximum eigenvalue tests to analyze the number of cointegrating vectors. The difference between the two test statistics is that the trace test is a joint test where the null hypothesis is that the number of cointegrating vectors is less than or equal to r, against a general alternative that it is more than r, whereas the maximum eigenvalue test conducts a separate test on the individual eigenvalues, where the null hypothesis is that the number of cointegrating vectors is r, against an alternative of (r + 1). The above equations are expressed as follows:

Trace test:

$$J_{trace}=-T{\sum }_{i=r+1}^n{1}_n(1-\widehat{{\lambda }_i})$$

(4)

The trace test examines the null hypothesis of r cointegrating vectors against the alternative hypothesis of n cointegrating vectors.

Eigenvalue test:

$$J_{max}=-T{1}_n(1-{\widehat{\lambda }}_{r+1})$$

(5)

The maximum eigenvalue tests the null hypothesis of r cointegrating vectors, whereas the alternative hypothesis states the existence of r + 1 vectors. Therefore, ${\lambda }_{r+1}\dots ,{\lambda }_n$ are the estimated eigenvalues of the trace and eigenvalue tests, respectively. $\lambda$ represents the estimated values from the estimated π matrix, and T represents the sample size.

3.3. Vector Error Correction Model

If a set of variables is found to have one or more cointegrating vectors, then a suitable estimation technique is a vector error correction model, which adjusts to both long-run and short-run changes in variables and deviations from equilibrium (Harris et al., 1995).

In this study, the VECM is expressed in Equation (6) as

$$\begin{array}{l}∆AE={\beta }_0+{\beta }_1{\sum }_{i=1}^{n-1}∆{ATM}_{t-i}+{\beta }_2{\sum }_{i=1}^{n-1}∆{RP}_{t=i}+{\beta }_3{\sum }_{i-1}^{n-1}∆{AL}_{t=i}+\\ {\beta }_4{\sum }_{i-1}^{n-1}∆{GDPPC}_{t=i}+\phi {ECT}_{t=1}+{\mu }_t\end{array}$$

(6)

where Δ is the first difference and the joint consequences of lags are ${\beta }_1$, ${\beta }_2$, ${\beta }_3$, and ${\beta }_4$, while the error correction term is specified by ${ECT}_{t=1}$ and φ is the adjustment speed of the model towards equilibrium. The error term is represented by $\mu$ and t represents the time period. The error term needs to be negative and statistically significant in order to bring about equilibrium. The study thereafter performed the diagnostic as outlined in the next section.

3.4. Residual Diagnostic and Stability Tests

The study employed serial correlation, a Lagrange multiplier, heteroscedasticity and the normality test to evaluate the reliability and validity of the results.

3.4.1. Lagrange Multiplier Test for Serial Correlation

When time across periods is correlated with the error term, serial correlation will occur (Mills, 2014). According to Rois et al. (2012), independent residuals are tested using the Breusch–Godfrey test, which tests serial correlation. The condition of serial correlation is that when the calculated Lagrange multiplier serial correlation is greater than 5%, then the null hypothesis is rejected, and vice versa.

3.4.2. Testing for Heteroscedasticity

The assumption of constant error variance in classical linear regression analysis is essential for producing dependable estimates. Heteroskedasticity testing is thus essential to verify if the variance of the error term is uniform (Breusch & Pagan, 1979). Therefore, white heteroscedasticity was applied as a potential remedy for the violation of this assumption. Hence, when the probability value exceeded the 5% level, suggesting the absence of heteroskedasticity, the null hypothesis was accepted. A probability below 5% level led to the rejection of the null hypothesis.

3.4.3. Test for Normality

The Jarque–Bera test statistic is defined as below:

$$JB={\displaystyle \frac{n}{6}}[S^2+{\displaystyle \frac{(K-{3)}^2}{4}}]$$

(7)

where n = sample size, K = sample kurtosis, and S = variable skewness. When the null hypothesis is above 5% probability, it is accepted; however, it is rejected below 5%.

3.5. Methodological Limitations and Future Research Extension

This study employed Johansen cointegration and the vector error model to analyze the impact of 4IR automation on agricultural employment in South Africa. These econometric techniques are well known for testing the long-run equilibrium and the short-run dynamics among non-stationary variables (Suykens, 2001). These methods are traditionally known to offer clear interpretation and hypothesis testing. Additionally, these methods enable the study to capture the short-run dynamics and the long-run equilibrium. However, these methods are not without limitations. For instance, they are limited in capturing non-linear cointegration and different technological regimes. Therefore, machine learning approaches such as artificial neural networks, support vector regression, the random forest model and gradient boosting could be employed to extend the analysis to capture complex non-linear patterns, and they offer predictive accuracy. Nevertheless, these machine learning approaches are constrained by their lack of necessary policy-driven interpretations.

4. Empirical Findings and Discussions

This section presents the results of the unit root, Johansen cointegration, vector error correction model and diagnostic tests.

4.1. Unit Root Results

In this study, the null hypothesis of non-stationarity is accepted if the ADF or PP calculated value exceeds the critical values at a 5% significance level and rejected if it is below (as seen in

Table 2. ADF and PP unit root test at 1st difference.

		Levels
Variables	Formula	ADF		PP		Remarks
		t-Statistics (Lags)	5% Critical Value	t-Statistics (Band Width)	5% Critical Value
AE	Intercept	−0.297	−2.957	0.836	−2.954	NSY
	Trend and intercept	−2.285	−3.558	−1.547	−3.553	NSY
	None	1.883	−1.952	4.930	−1.951	NSY
ATM	Intercept	−1.396	−2.954	−1.492	−2.954	NSY
	Trend and intercept	3.389	−3.389	−3.389	−3.553	NSY
	None	1.414	−1.952	1.837	−1.951	NSY
RP	Intercept	−1.199	−2.954	−1.213	−2.954	NSY
	Trend and intercept	−2.075	−3.603	−2.027	−3.553	NSY
	None	0.473	−1.951	0.656	−1.951	NSY
AL	Intercept	−1.584	−2.960	−2.332	−2.954	NSY
	Trend and intercept	−1.364	−3.563	−2.098	−3.553	NSY
	None	0.089	−1.952	−0.469	−1.951	NSY
GDPPC	Intercept	−0.124	−2.957	0.036	−2.954	NSY
	Trend and intercept	−2.013	−3.558	−1.785	−3.553	NSY
	None	1.689	−1.952	2.192	−1.951	NSY

NSY means non-stationary.

). If the series is non-stationary at level, it must be differenced until it becomes stationary to address the unit root issue.

The ADF and PP unit root tests at level revealed similar conclusions that AE, ATM, RP, AL and GDPPC appear to be non-stationary in all formulas. The results indicate that all the variables included are not stationary at level, necessitating differencing of the series as shown in

Table 3. Tests for augmented Dickey–Fuller (ADF) and Phillip–Perron (PP) unit root at 1st difference.

Variables	Formula	1st Difference
		ADF		PP		Remarks
		t-Values	5% Critical Value	t-Values	5% Critical Value
AE	Intercept	−5.764 **	−2.960	−5.778 **	−2.960	STN
	Trend and intercept	−5.662 **	−3.563	−5.670 **	−3.563	STN
	None	−5.863 **	−1.952	−5.880 **	−1.952	STN
ATM	Intercept	−6.353 **	−2.960	−7.326 **	−2.957	STN
	Trend and intercept	−6.279 **	−3.563	−7.253 **	−3.558	STN
	None	−6.060 **	−1.952	−6.340 **	−1.952	STN
RP	Intercept	−5.581 **	−2.654	−5.664 **	−2.957	STN
	Trend and intercept	−5.470 **	−3.558	−5.539 **	−3.558	STN
	None	−5.564 **	−1.951	−5.638 **	−1.952	STN
AL	Intercept	−7.528 **	−2.960	−6.017 **	−2.957	STN
	Trend and intercept	−7.715 **	−3.563	−6.474 **	−3.558	STN
	None	−7.619 **	−1.952	−6.154 **	−1.952	STN
GDPPC	Intercept	−3.740 **	−3.654	−3.322 **	−2.957	STN
	Trend and intercept	−3.571 **	−3.558	−3.573 **	−3.558	STN
	None	−2.792 **	−1.952	−2.693 **	−1.952	STN

Note: ** indicate significance levels at 5%. STN means stationary.

.

The ADF and PP unit root tests revealed similar conclusions that AE, ATM, RP, AL and GDPPC appear to be non-stationary (NSY). Thus, the results of the analysis revealed that all the variables are non-stationary at level, suggesting a further analysis is necessary to difference the series, as in Table 3.

The variables were then first-differenced as presented in Table 3, to reject the null hypothesis that indicates that the series has a unit root. After first-differencing all variables, agricultural employment, automation, rural population, arable land, and GDP per capita became stationary (STN), concluding that the variables were integrated in order I (1).

4.2. Lag Order Selection Criteria

The study used the information criteria approach to carefully select the lag length order to be used for this study. The lag length below was selected for the model.

The lag length selection criterion for the model is presented in

Table 4. Lag length selection results.

Lag	LogL	LR	FPE	AIC	SC	HQ
0	−1.427041	NA	6.41 × 10−8	0.464190	0.739016	0.555287
1	183.7709	289.3717 *	5.97 × 10−12 *	−8.860680 *	−6.936901 *	−8.223002 *
2	215.1657	37.28134	1.03 × 10−11	−8.572855	−5.000124	−7.388596

* Indicates lag order selected by the criterion. LR: sequential modified LR test statistic (each test at 5% level), FPE: final prediction error, AIC: Akaike information criterion, SC: Schwarz information criterion, and HQ: Hanna–Quinn information criterion.

. The results revealed that lag 1 is favorable according to the criteria. The study employed the Akaike criterion as it makes it possible to intervene when introducing one or more explanatory variables, without the loss of degrees of freedom. Therefore, the appropriate lag 1 was the most suitable to test Johansen cointegration.

4.3. Johansen Cointegration Results

Since the lag length had been acknowledged, it was important to test the long-run relationship that existed among the variables.

Table 5. Johansen cointegration test results.

Hypothesis Number of CE	Eigenvalue	Trace Statistics	0.05 Critical Values	Max-Eigen Statistics	0.05 Critical Values
None *	0.720347	98.05397	95.75366	40.77458	40.07757
At most 1	0.414000	57.27939	69.81889	17.10191	33.87687
At most 2	0.398489	40.17748	47.85613	16.26596	27.58434
At most 3	0.327728	23.91152	29.79707	12.70696	21.13162
At most 4	0.252331	11.20456	15.49471	9.305429	14.26460
At most 5	0.057621	1.899135	3.841466	1.899135	3.841466

Trace test indicates 1 cointegrating eqn at the 0.05 level. Max-eigenvalue test indicates 1 cointegrating eqn at the 0.05 level. * Denotes rejection of the hypothesis at the 0.05 level.

below presents the results of Johansen cointegration.

According to the results provided in Table 5, the trace test and maximum eigenvalue test show one cointegrating equation. Under the trace test, the null hypothesis is rejected at None, where the trace statistic of 98.05397 is greater than the 5% critical value of 95.75366, confirming cointegration. The maximum eigenvalue test also indicates one cointegrating equation at None with a maximum-eigenvalue statistic of 40.77458, which is greater than the 5% critical value of 40.07757. From At most 1 to At most 5 for both the trace test and maximum eigenvalue test, the null hypothesis is not rejected as the trace and maximum eigen statistics are less than the 5% critical values. Therefore, since cointegration was found, VECM was performed in the following section.

4.4. Vector Error Correction Model Results (VECM)

The existence of cointegration between variables suggests that there is a long-run relationship that exists between the dependent variable and the independent variables.

Table 6. Estimated long-run parameters.

CointEq1	Coefficients	Standard Errors	t-Values	Conclusion
Constant		-	-	-
Variables
AE	1.000000	-	-	-
ATM	0.288	0.079	3.648	Negative and Significant
RP	−0.350	0.074	−4.704	Positive and Significant
AL	−0.694	0.061	−11.278	Positive and Significant
GDPPC	0.818	0.030	27.124	Negative and Significant

below presents the results of the long-run parameters, followed by the short-run parameters in

Table 7. Short-run parameter results.

	Coefficient	Standard Error	t-Statistics	Conclusion
ECT	−0.188	0.086	−2.187	Negative and significant
Variables
D(ATM(-1))	0.019	0.050	0.386	Negative and Insignificant
D(RP(-1))	−0.073	0.046	−1.582	Positive and Insignificant
D(AL(-1))	0.037	0.096	0.390	Negative and Insignificant
D(GDPPC(-1))	−0.201	0.257	−0.780	Positive and Insignificant
R2	0.670	-	-	67% variation is corrected by the independent variables
Adj R2	0.572	-	-	Model is good fit
F-statistics	6.930	-	-	Significant model

.

The long-run results presented in Table 6 revealed that a one percent increase in automation leads to a 29% decrease in agricultural employment, as shown by the negative and significant relationship. This implies that as farmers use more machines, many low-skilled farm workers may lose jobs, which can increase rural unemployment and poverty if new jobs are not created in other sectors. This supports the findings of Yang and Li (2023) and Apicella (2025) in their studies.

However, rural population was revealed to have a positive and significant relationship with agricultural employment; this indicates that a one percent increase in rural population increases agricultural employment by almost 35% in the long run, proving that more people living in rural areas depend on farming for jobs and income, as supported by Shen et al. (2023) in their study.

In addition, arable land was also proven to have a positive and significant relationship with agricultural employment in South Africa, indicating that a one percent increase in arable land increases agricultural employment by 69% percent in the long run. This highlights that increased availability of arable land signals an increase in agricultural activities, which may require more people to be employed for agricultural activities. This is in line with what Xiao and Zhao (2020) found, as their study indicated that the higher the availability of land, the lower the probability of working in a non-agricultural sector.

Lastly, a one percent increase in GDP per capita decreases agricultural employment by 81% in the long run as shown by a negative and significant relationship. This reflects structural transformation of workers moving out of low-productivity farming into higher-productivity sectors such as services and industries as incomes grow (Roser, 2023; Kabini, 2022).

Table 7 results revealed that the coefficient of the error correction term is negative (−0.188) and statistically significant with a t-statistic of −2.187. This implies that it will take a speed of 19% for the systems of agricultural employment to adjust to equilibrium in a single year. All the variables are statistically insignificant in the short run, indicating that automation, rural population, arable land, and GDP per capita do not have a significant relationship with agricultural employment.

4.5. Diagnostic Test Results

The following results in

Table 8. Diagnostic test results.

Test	H0	Probability	Conclusion
Serial Correlation	No serial correlation	0.6014	No serial correlation
White (CH-sq)	No heteroskedasticity	0.1882	No heteroskedasticity
Jarque–Bera	Residuals are normally distributed	0.1907	Residuals are normally distributed
Reject if p < 0.05

outlines the diagnostic procedures performed on the model, specifically tests for serial correlation, heteroskedasticity and normality. The results indicate no evidence of serial correlation and heteroskedasticity. The residues are also normally distributed. This suggests that the model is well specified and the coefficients are reliable for statistical inference.

5. Conclusions and Recommendations

The study has shown that the 4IR is restructuring the agricultural sector within South Africa. Studies have indicated the benefits of automation on agriculture, which include efficiency and accuracy of data. However, this comes at the cost of human labor as manual jobs become replaced by machines. The findings of the study show that automation, rural population, arable land, and GDP per capita do not have a significant relationship with agricultural employment. The results indicate an increase in unemployment due to the introduction of the 4IR within the agricultural sector. Consequently, policy interventions should be tailored to farm type by prioritizing precision agriculture among small-scale farmers and advanced automation in large-scale commercial farms that benefit from economies of scale and capital capacity. Conversely, small-scale/subsistence farming requires automation that is affordable but improves productivity without undermining rural employment. Therefore, policies should focus on cost-effective technologies that can lower production risk, enhance input and improve farm income while preserving employment. Additionally, the study recommends skills training for agricultural employees to keep up with advanced technologies and also recommends governmental support on programs that will protect vulnerable employees from being replaced by 4IR machines.