The performance of machine learning models depends on the choice of hyper-parameters. Moreover, in many cases one would train different machine learning models (e.g. SVMs or Gradient Boosting), and choose the better performing one.
Instead of manually trying out hyper-parameters or performing manual model selection, one can automate these tasks and gain experimentation speed and performance.
However, it is still recommended that models are peer-reviewed and assessed by team members before deployment to production.