Two tools reps actually use: three inputs become a phased proposal draft (live LLM), and a deal-risk board scores pipeline with deterministic rules — days since reply, single-threading, pushed close dates — where every point traces to a visible flag.
The risk model is intentionally rules-first: reps trust '19 days silent + single-threaded + pushed twice' far more than an opaque 0.73. AI's job here is the follow-up draft, not the verdict.
