People mostly spend their time indoors for their daily activities. However, indoor air pollutant concentrations are found to be higher than outdoors. Generally, this is caused by the ventilation performance that is not able to dilute indoor air pollutants adequately. The presence of indoor CO2 at certain concentration level is an indicator of indoor air quality and requires field measurements to evaluate it. On the other hand, the consequence of field measurements is not only time consuming but also costly. In order to minimize that problems, this study aimed to predict the model of indoor air quality by referring to previous data. It was achieved by several stages such as input, process, and output. A number of previous data regarding indoor air quality namely indoor CO2, indoor temperature, number of occupants, and air conditioner usage duration were assigned as input. Subsequently, the process stage in this study adopted feed-forward neural networks that divided the data into training data and testing data. Additionally, several activation functions in neural network such as ReLU, tanh, logistic, and identity were involved in the process phase in order to imitate the actual model precisely. Ultimately, the outputs were evaluated using mean square error, mean absolute percentage error, and coefficient of determination. The findings indicated that the application of logistic as activation function was prominently reliable to predict the targeted data. This activation function can improve learning performance which is characterized by their value of mean square error, mean absolute percentage error, and coefficient of determination. In addition, a number of discrepancies of each activation functions were also presented to identify their behavior in terms of imitating the given data. Finally, this approach can be used as a tool to predict the concentration level of indoor CO2 in a concise time and leads to cost efficiency.