The quarter's growth target — 250 new customers — was mathematically impossible at current onboarding cost. So opex became the growth lever: cutting 65% of it turned an impossible plan into a hit quarter. AI didn't decorate the product here; it unblocked the revenue plan.
The reality behind that math: 2–3 implementation managers hand-tagging ~100 lease documents per customer in Adobe PDF — days to weeks per customer. No pod had capacity to fix it.
So I changed the math. I partnered with a floating AI engineer outside my pod and built the front end myself in Replit: upload a blank lease, AI identifies and places the tags the way a human would, implementations reviews instead of typing. Nearly free to build — no roadmap capacity consumed.
Result: 65% opex saved on onboarding, documents tagged and validated in days instead of weeks, and the team hit the 250-customer objective because of it.
Product signals lived in silos — behavioral analytics, support tickets, ops data — and synthesizing them manually ate days every month. I built an internal production tool on Claude + Vercel that correlates behavioral, operational, and support signals and surfaces actionable insights. Prototyped it, proved value, productionized it. It became part of how the product team makes decisions.
AI adds leverage where judgment is cheap and volume is expensive: synthesis, drafting, routing, tagging. It's a risk where trust is the product — compliance-sensitive flows, money movement, anything a resident signs. I've also argued against AI when plumbing was the answer: the lease-rendering fix that moved CSAT from 6 to 9 was a file-format migration, not a model.
The tooling matters less than the habit: prototype it yourself, prove the value, then productionize. The context you give a model is a product surface — I treat prompt and tool design the way I treat UX.