Depending on how you are doing your deduplication, this could be due to your deduplicated reads being concentrated in your BED features. For example, if this were RNAseq, you find, that on average only 1/2 to 2/3 of your reads will map to annotated transcripted regions, with the rest being intronic reads, transcript-noise, DNA contamination etc. The signal is much higher in genes, but the non-geneic reads still account for a large fraction of the whole. If reads in genes are more highly duplicated then you will see a bigger change in the genic regions than outside them.

Also remember that if your data is pair, UMI-Tools will report the number of pairs input, but if your pairs are overlapping, then they will contribute 2 to the coverage.