The aim of this study was to investigate whether Timed Up-and-Go (TUG) dual-task (TUGdt) tests predict dementia incidence among patients with subjective or mild cognitive impairment (SCI; MCI). Other study objectives were to determine whether TUGdt improves dementia prediction compared to a) demographic characteristics and standard cognitive tests alone; and b) TUG and Verbal Fluency performed separately. Patients (n = 172, age range 39–91 years, 78 women) with SCI or MCI performed TUGdt tests, including 1) naming animals and 2) reciting months backwards, and clinical cognitive tests at baseline. Diagnoses were identified at follow-up after 2.5 years. Logistic regression was used to predict dementia incidence, receiver operating characteristic (ROC) curves and c-statistics for predictive capacity. Analyses were stratified by age and gender. At follow-up, 51 patients had developed dementia. The TUGdt result “animals/10 s” was associated with dementia incidence (standardized odds ratio (OR) = 4.06, 95% confidence interval (CI) 2.28–7.23, p < 0.001), more so among patients under the median age of 72 years (standardized OR = 19.4, 95% CI 3.53–106.17, p < 0.001). TUGdt “animals/10 s” improved dementia prediction compared to demographic characteristics and standard tests alone (c-statistics 0.88 to 0.94) and single-task tests (c-statistics 0.86 to 0.89), but only in the younger patient group. TUGdt has the potential to become a useful tool for dementia prediction.