Abstract: Low groundwater levels harm groundwater-dependent ecosystems, agriculture and infrastructure. Managers therefore need the probability that heads fall below critical levels, not just a simulation of past heads. This study combined transfer function-noise (TFN) models with Monte Carlo simulation to assess the reliability of monthly groundwater levels at three shallow observation wells in the Netherlands (B58C0698 near Heibloem; B32C0639 and B32C0609 near De Bilt). Pastas TFN models with a gamma response function, linear recharge and AR(1) noise explained 71–93% of the head variance. Five hundred synthetic 30-year sequences were generated for each well by block-bootstrapping hydrological years of observed daily precipitation and evaporation. Parameter uncertainty and residual noise were added stepwise, and a dry stress scenario (precipitation −5%, evaporation +10%) and uniform drawdowns of up to 0.5 m were also simulated. Failure was defined as a monthly mean head below the 10th percentile of observed monthly heads (Q10); Q05 and the mean lowest groundwater level (GLG) were used as alternative thresholds. With all uncertainty sources included, time-based reliability was 0.89–0.90 and the annual probability of at least one failure month was 0.35–0.50, consistent with the observed frequencies. Parameter uncertainty had a negligible effect, whereas residual noise reduced reliability by up to 0.03 and increased the annual failure probability by up to 0.09. Failures were concentrated in August. Resilience was low (0.33–0.43), and failure events lasted 4.4–7.1 months at most on average. The dry stress scenario reduced reliability to 0.77–0.83 and raised the annual failure probability to 0.53–0.78. A uniform drawdown of only 3–11 cm lowered reliability below 0.85, and the De Bilt well with the smallest head variability was the most sensitive. TFN-based Monte Carlo analysis is a practical tool for setting risk-based limits on groundwater abstraction.