BACKGROUND: Praxis impairment may be one of the first symptoms manifested in dementia, primarily in cortical dementia. The Cambridge Cognitive Examination (CAMCOG) evaluates praxis, but little is known about the accuracy of CAMCOG for diagnosing dementia. The aims here were to investigate the accuracy of praxis and its subitems in CAMCOG (constructive, ideomotor and ideational subitems) for diagnosing Alzheimer's disease (AD) among elderly patients. DESIGN AND SETTING: Cross-sectional study on community-dwelling elderly people. METHODS: 158 elderly patients were evaluated. CAMCOG, Mini-Mental State Examination and Pfeffer Functional Activities Questionnaire were used. ROC curve analysis was used to establish cutoff points. RESULTS: The total scores for praxis and the constructive subitem presented significant differences (P < 0.0001) between healthy elderly people and AD patients. Stage of dementia (clinical dementia rating, CDR = 0, 1 and 2) showed that total and constructive praxis can be used to classify the stages of dementia (mild and moderate cases), i.e. constructive praxis classified 88% of the patients with mild dementia (P < 0.0001) while total praxis classified 56% with moderate dementia. Comparison of normal controls (NC) and mild dementia cases showed specificity of 71% and sensitivity of 88% (AUC = 0.88; P < 0.0001). CONCLUSION: Some praxis subtests can have higher predictive diagnostic value for detecting Alzheimer's disease in mild stages (total praxis AUC = 0.858; P < 0.0001; constructive AUC = 0.972; P < 0.0001). Constructive praxis as measured using CAMCOG may contribute towards diagnosing dementia, because occurrence of impairment of praxis may help in recognizing an evolving dementia syndrome.