<h4>Background: </h4> Congenital Adrenal Hyperplasia (CAH) due to 21-hydroxylase deficiency is caused by biallelic pathogenic variants in CYP21A2. Although well-established genotype-phenotype correlations exist for the common pseudogene-derived alleles, next-generation sequencing increasingly identifies rare missense variants whose clinical significance is unknown, which hampers classification, counselling and the selection of patients for new targeted therapies. <h4>Methods:</h4> We characterized three CYP21A2 missense variants reported in patients with non-classic (NC) or simple-virilizing (SV) CAH, p.A90V (c.269C&gt;T), p.L108Q (c.323T&gt;A) and p.L318M (c.952C&gt;A), together with the intron 2 splice variant c.293-13C&gt;G (I2G) found in trans in two of the patients. Variants were annotated at the gene level with NCBI E-utilities (ClinVar, dbSNP), Ensembl VEP, SpliceAI and AlphaGenome, at the population level with gnomAD v4.1 and the 1000 Genomes Project, and at the protein level with AlphaMissense, REVEL, CADD, MetaRNN, PrimateAI, evolutionary conservation, AlphaFold and FoldX stability calculations mapped on the CYP21A2 crystal structure. Wild-type and variant enzymes were expressed in HEK293T cells, protein levels were quantified by Western blot, and 21-hydroxylase activity was measured with radiolabelled progesterone and 17-hydroxyprogesterone (17-OHP) followed by thin-layer chromatography. <h4>Results:</h4> All three missense variants are absent from or singletons in 1.6 million gnomAD alleles, absent from the 1000 Genomes Project and absent from ClinVar, affect residues that are conserved in mammalian orthologues and lie in high-confidence regions of the AlphaFold model (pLDDT 97-99). AlphaMissense classified p.A90V and p.L108Q as likely pathogenic (0.76 and 0.85) but p.L318M as likely benign (0.31), whereas CADD, MetaRNN and REVEL supported deleteriousness for all three; SpliceAI and AlphaGenome excluded a splicing effect of the missense changes but reproduced the acceptor-site loss and cryptic-acceptor activation caused by I2G. In vitro, protein expression was reduced to 52%, 65% and 59% of wild type for p.A90V, p.L108Q and p.L318M. Total 17-OHP 21-hydroxylation was 36.7%, 1.4% and 38.5% of wild type, and total progesterone conversion 24.9%, 9.2% and 17.5%, revealing a substrate-dependent, almost complete loss of function for p.L108Q. FoldX predicted destabilization for p.A90V (ΔΔG 3.2 kcal/mol) and p.L108Q (2.5 kcal/mol) but not for p.L318M (0.1 kcal/mol); structural analysis attributes the severe p.L108Q defect to disruption of the hydrogen-bond network (L108-G111-R427) that anchors the heme propionates. <h4>Conclusions:</h4> Combining functional assays with gene- and protein-level computational analysis classifies all three variants as pathogenic. p.L108Q is a severe allele consistent with classic SV CAH, which resolves the ambiguous SV/NC classification of the male index patient, whereas p.A90V (homozygous) and p.L318M (in trans with I2G) are mild alleles that explain the NC phenotypes. The data illustrate that AlphaMissense alone would have misclassified p.L318M and that residual activity, not in silico scores, should guide counselling and the selection of patients for CRF1-receptor-antagonist therapy.