The aim of this study was to examine predictions derived from a proposal about the relation between word-decoding skill and working memory capacity, called verbal efficiency theory. The theory states that poor word representations and slow decoding processes consume resources in working memory that would otherwise be used to execute high-level comprehension processes, such as the generation of inferences. Previous research has yielded inconsistent findings about the importance of word decoding in adult readers, and the hypothesis has never been tested experimentally. Verbal efficiency theory was tested in this experiment by manipulating the difficulty of grapheme-phoneme conversion and assessing the extent to which readers made bridging inferences. Participants read two-sentence passages and then responded to lexical decision targets. Some of the passages required a bridging inference to integrate the first and second sentences. Decoding difficulty was manipulated such that the second sentence in some passages was written using pseudohomophones. Participants also received tasks to assess their working memory capacity and decoding ability. Inference priming was found in both the Standard American English and pseudohomophone contexts but was stronger in the former than in the latter. The advantage in priming for the Standard English relative to the pseudohomophone condition was predicted by an interaction between decoding skill and working memory capacity. Poor decoders who scored high on the span tests were less impaired by the pseudohomophone manipulation than were poor decoders who scored low on the tests. The results suggest that working memory capacity compensates for poor decoding skills even among proficient adult readers.