You answered your own question. You can get rid of the manual review. Plus, if you can just point a big-ass data stream at Watson and it can actually ferret out malfeasance, you can also get rid of the folks who program those "other forms of automation".
Instead, you'll be replacing them with a smaller number of people who can choose training sets and interpret Watson's output. You'll also see savings in your programming costs, as you've replaced several fraud detection systems with a single, unified system. So you get a smaller workforce.
You also get a more bifurcated workforce, with a small number of jobs being up-skilled (and more-highly paid), while a larger number of jobs are down-skilled or eliminated altogether. Whether this works when translated to a large portion of the white-collar, service-sector workforce as a whole is left as an exercise to the reader.