I think it's safe to say that deductive reasoning is usually taken to describe types of inference where the truth of the premises guarantees the truth of the conclusion in a robust, or stable way, i.e., in such a way that this doesn't break down upon acquiring new information (provided also, as Henning mentions in his answer, that terms are well defined, unambiguous, the rules of deduction are properly used, etc.).
More concretely, if a conclusion $\varphi$ follows from a set of premises $\Gamma$, then $\varphi$ should also follow if we add more stuff to $\Gamma$. In other words, the truth of $\varphi$ is firmly established by the truth of $\Gamma$, and there's no fear that we'll have to take back $\varphi$ if we learn new things (i.e., the inference relation is assumed to be monotonic).
Falling outside the scope of this category there would be types of inference where the truth of the premises by themselves is, in some way, not enough to guarantee the truth of the conclusion. These would be cases where the truth of the premises $\Gamma$ supports the truth of the conclusion $\varphi$ only up to a certain degree (as in probabilistic reasoning); only in 'typical' (or most) circumstances; or only under special assumptions, e.g., that there is no evidence to the contrary. These are cases where we're likely to change our assessment of $\varphi$ as more information comes in.
Many of these patterns of inference can be found in common-sense reasoning: the type of 'imperfect' reasoning humans engage in on a day-to-day basis to navigate the world. And many of these are of interest to computer scientists and AI researchers, and are studied formally under the heading of defeasible reasoning and non-monotonic logic (links are to the Stanford Encyclopedia of Philosophy articles on the topics).