Automatic email answering is a difficult AI problem that combines classification, natural language understanding and text generation techniques. We present an original approach and a tool based on an ontology to automatically reply to meeting emails. We constructed the ontology from a French corpus of 1150 emails in which the concepts represent detailed meeting intents (proposing a meeting, cancelling a meeting, rescheduling a meeting) and different answer templates. Each intent concept is a semantic rule formalized according to the FrameNet methodology. These rules are used to detect intents in emails and also to extract relevant information (such as date, time or person) used for generating replies. The main advantage of our approach is the generation of more precise answers than those proposed by other approaches. We tested the intent detection step on a set of 297 emails and compared it with different supervised machine learning algorithms. Obtained results are encouraging, with an accuracy 20% higher than results obtained with other algorithms. keywords : Ontology engineering, knowledge acquisition from text, knowledge-based recommendation systems.