To describe the therapeutic drug monitoring (TDM) of cefepime in non-critically ill adults and compare four different ways of dosing: conventional table-based; empirically adjusted following TDM; individualized based on a population pharmacokinetic (PopPK) model without TDM; and TDM-adjusted with a Bayesian approach integrating TDM and PopPK. We conducted a retrospective study in a tertiary centre to examine the current practice of TDM and to evaluate the potential for improvement by PopPK-based software individualization. The prediction of trough concentrations and the total daily doses (TDD) prescribed according to each approach were compared by calculating the mean logarithmic bias and the root mean squared error, complemented by linear regression and variance analysis. Among 168 trough concentrations in 119 patients (median: 12 mg/L), 38.6% of measurements exceeded 15 mg/L, the reported threshold for neurotoxicity. Nine patients developed neurotoxicity. The prediction performance of PopPK alone for trough concentrations was moderate, but clearly improved after integration of TDM. Accordingly, TDD were significantly lower for a priori PopPK-based dosage (mean: 2907 mg/24 h) compared with actual table-based dosage (4625 mg/24 h, P < 0.001). They were also lower for a posteriori dosage based on PopPK and TDM (3377 mg/24 h) compared with actual dosage after empirical TDM (4233 mg/24 h, P < 0.001), as model-based adjustment privileged more frequent administrations. Our observations support routine TDM of cefepime to prevent overdosing and subsequent toxicity in the non-critically ill. Software-based individualization seems promising to optimize the benefits of TDM, but has little potential to replace it.