The deposition prep workflow that finally made me trust the agents

I'm a paralegal at a 14-attorney litigation boutique. For 18 months I've been pushing AI into our workflow and most of it was theater until I rebuilt our deposition prep stack around three narrow agents instead of one big one. The pattern that broke: a single Claude project handling exhibit indexing, timeline construction, and witness inconsistency flagging. It hallucinated dates roughly 1 in 30 documents. Partners stopped trusting it within a month. What works now: separate agents per task, each with a structured output schema and a human-readable confidence column. Exhibit indexing runs first, output is locked, then timeline ingests only that. Inconsistency flagging runs last and cites Bates numbers I can click through. My prep time on a mid-size case dropped from about 22 hours to 7. But the real change is that the associates actually read the output now because it's not pretending to be a brief, it's pretending to be a junior paralegal's worksheet. The productivity story isn't models getting smarter. It's me getting better at scoping them.