Disease modifying drugs for Alzheimer's disease (AD) are likely to be most effective when given in non-demented subjects. In this review we summarized biomarkers in cerebrospinal fluid (CSF) and blood that can predict AD-type dementia in subjects with mild cognitive impairment (MCI). In addition, we investigated whether these markers could reduce sample size and costs if used to select subjects for trials on the prevention of AD in subjects with MCI. A meta-analysis of markers that had been investigated in multiple studies showed that the combination of amyloid-beta (Abeta1-42 and tau in CSF had the best predictive accuracy for AD (odds ratio (OR) 18.1, 95% confidence interval (CI) 9.6-32.4). Abeta1-42, total tau, and phosphorylated tau in CSF also predicted conversion, but with lower accuracy (OR 7.5 to 8.1). Plasma levels of Abeta1-40, Abeta1-42, the ratio Abeta1-42/Abeta1-40 and homocysteine did not predict outcome. In a fictive trial design, the use of the combination of Abeta1-42 and tau in CSF in the selection of subjects could reduce sample size by 67% and trial costs by 60% compared to a trial in which unselected subjects with MCI would be enrolled. In conclusion, the combination of Abeta1-42 and tau in CSF is useful to select subjects for trials that aim to slow down the progression from MCI to AD-type dementia.