The challenge
The qualification rubric was tribal knowledge, partly written down, partly in the head of the lead partner. Misroutes were costly because specialists had narrow practices. The firm had tried two off-the-shelf "AI lead qualification" SaaS tools and scrapped both because neither learned the rubric correctly and both kept misclassifying borderline cases.
What we did
Two-week strategy sprint, then a 21-day pilot. Wrote the rubric properly with the lead partner. Built a Claude-powered agent that reads the inbound form, enriches against the firm's research stack, scores against the rubric, and routes to the right partner. Low-confidence cases get routed to a review queue instead of being guessed at.
The outcome
Live in 19 days, two days ahead of plan. Replaced 18 hours per week of partner time. Agreed with the lead partner on 94% of confident cases. Zero misroutes after the eval bar was set, because the agent flags uncertainty instead of guessing.