A common "rule of thumb" is that ordinal Likert-type items with 5 or more categories can potentially be treated like continuous indicators without running into too much trouble (i.e., using covariances & maximum likelihood [ML] estimation in CFA/SEM), provided they have rather symmetric item distributions. If the item distributions are heavily skewed, you might run into trouble because using item covariances as input and ML estimation (assuming multivariate normality!) may be inappropriate in that case.

With skewed distributions or fewer than 5 item categories, use of polychoric correlations and WLSMV estimation (or at least using robust ML estimation) is recommended. I'm not sure whether this is possible in AMOS. It looks like AMOS offers Bayesian estimation with ordinal data, but I don't know how well this works compared to WLSMV or robust ML estimation with ordinal data.

See: Chapter Non-normal and categorical data in structural equation modeling