We outline SEEXC, a neural network model of choice response time (RT) based on leaky competitive integrators. SEEXC is different from extant neutral network models in that it incorporates the effects of practice by modulating recurrent selfconnection weights. For simplified versions of this model, we provide analytic and numeric results concerning RTs and the relationship between RT and practice – the “Law of Practice” – that match those observed empirically. We also show that previous methods of modelling practice in similar systems, which modulate inputs, are unlikely to successfully match observed data. The simplified versions of the model analysed are appropriate for modelling the non-stochastic parts of simple RT and two-choice RT, provide insight into the behaviour of the full version of SEEXC, and suggest a new form for the Law of Practice.