Background and Aims: We hypothesized that artificial intelligence (AI) models are more precise than standard models for predicting outcomes in acute-on-chronic liver failure (ACLF). Methods: We recruited ACLF patients between 2009 and 2020 from APASL-ACLF Research Consortium (AARC). Their clinical data, investigations and organ involvement were serially noted for 90-days and utilized for AI modelling. Data were split randomly into train and validation sets. Multiple AI models, MELD and AARC-Model, were created/optimized on train set. Outcome prediction abilities were evaluated on validation sets through area under the curve (AUC), accuracy, sensitivity, specificity and class precision. Results: Among 2481 ACLF patients, 1501 in train set and 980 in validation set, the extreme gradient boost-cross-validated model (XGB-CV) demonstrated the highest AUC in train (0.999), validation (0.907) and overall sets (0.976) for predicting 30-day outcomes. The AUC and accuracy of the XGB-CV model (%Delta) were 7.0% and 6.9% higher than the standard day-7 AARC model (p < .001) and 12.8% and 10.6% higher than the day 7 MELD for 30-day predictions in validation set (p < .001). The XGB model had the highest AUC for 7- and 90-day predictions as well (p < .001). Day-7 creatinine, international normalized ratio (INR), circulatory failure, leucocyte count and day-4 sepsis were top features determining the 30-day outcomes. A simple decision tree incorporating creatinine, INR and circulatory failure was able to classify patients into high (similar to 90%), intermediate (similar to 60%) and low risk (similar to 20%) of mortality. A web-based AARC-AI model was developed and validated twice with optimal performance for 30-day predictions. Conclusions: The performance of the AARC-AI model exceeds the standard models for outcome predictions in ACLF. An AI-based decision tree can reliably undertake severity-based stratification of patients for timely interventions.