The contract negotiations lawyers deal with every day turn out to be more repetitive than they look. Uber's legal team processes thousands of contract negotiations a year, and simply responding to client edits round after round had become a serious bottleneck. But buried inside that high-stakes work was something surprisingly structured: similar clauses, repeatable redlines across clients, and standing legal reasoning patterns. That meant the routine parts could be automated while complex judgment stayed with the lawyers. 🔄 The first attempt used retrieval-augmented generation, but it failed — semantic matching between search keys and content was inaccurate, the tone of generated responses was off, and it generalized poorly to new scenarios. The team evolved it into a feedback loop that carefully logged every lawyer decision (accept, reject, modify) with exponential decay weighting to avoid drifting toward outdated precedent. Next they tamed responses that swung between overly defensive and overly accommodating using expert-defined prompt structure and a reflection technique, and finally reached agentic modifications capable of drafting counter-proposals themselves. Title: Scaling AI in Legal: Building Uber's Redlining Agent URL: Built right into the Microsoft Word lawyers already use, the tool cut contract review time by 20% and reached 91% accuracy on AI-generated decisions. What stands out most is the team's own takeaway: the hardest part wasn't the model — it was successfully weaving an agent into lawyers' daily work. #AIAgents# #LegalTech#