Making complex
systems simple.

A sprawling product surface meant support volume was scaling faster than the team could staff. The assistant had to understand intent across dozens of account states without sending users back to humans unnecessarily.

Mapped the top 40 intents against real conversation logs, then rebuilt the automation tree around what users were actually trying to do — not what the org chart said. Built a feedback loop so low-confidence turns went to human review and fed back into training.

Reached ~97% containment on covered intents while lifting conversion on paywalled flows. The same pattern now powers three adjacent surfaces.

Two channels with very different expectations — SMS needed to feel like a friendly text, web needed to feel like a polished concierge. Both had to hand off to a human without losing state.

Designed a channel-aware response layer on top of a single intent model, so the same brain produced channel-native copy. Instrumented every handoff so we could measure real containment, not just deflection.

Lifted automation success materially, reduced engineering overhead on maintenance, and unlocked $5M+ in annualized cost savings.