Time series are often “sticky” – a value will be correlated with the prior value, the one before that, and so on. The degree of correlation will typically decline the further back you go. AR models seek to use that auto-correlation information to produce forecasts.