Although it is well recognized that MCI represents a risk state for subsequent dementia, estimates of conversion vary widely according to the diagnostic criteria employed. There are currently no simple cognitive predictors of high and low risk of progression. We followed 107 non-demented non-depressed subjects from an original cohort of 124--sub-classified as follows: pure amnestic MCI (22), multi-domain MCI (54), non-amnestic MCI (10) and worried well (21). At 2 years, outcome varied considerably. Of the multi-domain MCI group 59% progressed to dementia and only 5% improved. By contrast, in pure amnestic MCI only 18% progressed and 41% improved. Of non-amnestic MCI patients 70% improved. The best predictor of progression was a combination of the Addenbrooke's cognitive examination (ACE) and the paired associate learning task (PAL), which produced high negative predictive (90%) and sensitivity (94%) values. The results indicate very different outcomes according to whether patients have pure amnestic versus multi-domain MCI. While the latter is an aggressive disorder, the former is more benign and unstable even in a clinic setting. Patients with scores >88 on the ACE and/or <14 errors on the PAL can be confidently reassured of a good prognosis.