I'm a paralegal at a mid-size litigation firm, 14 attorneys. Eighteen months ago we piloted an LLM-based doc review tool on a contract dispute with ~80k documents. The pitch was the usual: 70% time savings on first-pass review. What actually happened: first-pass went from six weeks to nine days. Great. But the second-pass QC ballooned. The model was confident on the wrong things, especially anything involving carve-outs, side letters, or oral modifications referenced in email threads. Associates stopped trusting the privilege flags entirely after one near-miss with a Rule 502 issue. So now I spend less time tagging and more time auditing the model's tags. My week looks roughly the same in hours, just shifted upstream of the attorneys instead of downstream. The firm books it as a win because partner hours dropped. Mine didn't. The honest version of the productivity story is: the cheap labor got cheaper, the expensive judgment got more expensive, and the middle layer (me) absorbed the variance. I don't think anyone modeling these ROI numbers is counting that.