Amyloid-beta (Aβ) 42/40 ratio, tau phosphorylated at threonine-181 (p-tau), and total-tau (t-tau) are considered core biomarkers for the diagnosis of Alzheimer’s disease (AD). The use of fully automated biomarker assays has been shown to reduce the intra- and inter-laboratory variability, which is a critical factor when defining cut-off values. The calculation of cut-off values is often influenced by the composition of AD and control groups. Indeed, the clinically defined AD group may include patients affected by other forms of dementia, while the control group is often very heterogeneous due to the inclusion of subjects diagnosed with other neurological diseases (OND). In this context, unsupervised machine learning approaches may overcome these issues providing unbiased cut-off values and data-driven patient stratification according to the sole distribution of biomarkers. In this work, we took advantage of the reproducibility of automated determination of the CSF core AD biomarkers to compare two large cohorts of patients diagnosed with different neurological disorders and enrolled in two centers with established expertise in AD biomarkers. We applied an unsupervised Gaussian mixture model clustering algorithm and found that our large series of patients could be classified in six clusters according to their CSF biomarker profile, some presenting a typical AD-like profile and some a non-AD profile. By considering the frequencies of clinically defined OND and AD subjects in clusters, we subsequently computed cluster-based cut-off values for Aβ42/Aβ40, p-tau, and t-tau. This approach promises to be useful for large-scale biomarker studies aimed at providing efficient biochemical phenotyping of neurological diseases.