One of the important variables in chemical process industries is the liquid level. The challenges of estimating the third-order liquid level system are: the system exhibits non-linear behavior, the system has coupled states where changes in one tank directly affect others, uncertainty where there are unknown disturbances, and prediction complexity. Yet, with advances in artificial intelligence (AI), modeling and estimating the liquid third-order levels has become a reality. The present work attempts to construct an accurate AI model to estimate the levels of third-order liquid level tanks utilizing the available inlet liquid flow rate and the bottom level. In order to estimate the dynamic level responses as a function of time for three liquid level tanks connected in series, this article presents the training of several types of artificial neural networks (ANNs): basic ANNs with different settings, autoformer, transformer regression, and attentive interpretable tabular learning (TabNet). These ANNs are trained to estimate the two unmeasured top and middle levels using the simulated response data of the bottom tank's level (measured level) for different step changes in the inlet water feed flow rate ranges from 0.1 to 3 liters/min. The findings, using the coefficient of determination (R^2), demonstrated that autoformer produces the best closed-fit responses; R^2 = 99.77% for the top-level estimation and R^2 = 99.86% for the middle-level estimation. Advanced AI techniques act as highly computationally efficient surrogate models for conventional complex models based on nonlinear conservation without the need for heavy computing power for process control and optimization.