Intercom

Intercom runs 500 people in R&D against 8.5 million lines of code, 313 production deploys a day, and 30,000 B2B customers. Nine months after «Darragh Curran» publicly committed to doubling R&D productivity in twelve months, he published a report card on 16 April 2026 claiming 3× over sixteen months with no plateau in sight.

The numbers are the reason this case keeps coming back. 93.6% of pull requests are agent-driven, 19.2% are AI-approved against a stated target above 50%, defect backlog is down 54%, cost per PR down 50%, downtime from breaking changes down 35%. Median merge time for «Claude Code» in auto-approved mode is 14.6 minutes, against 75.8 minutes for the organisation as a whole. Curran treats the merged PR as the pivot metric and describes R&D as a factory producing high-quality increments.

The organisational piece is the «Skills-Based Plugin Architecture», a private marketplace distributing specialised «Claude Code» configurations with auto-update: 153 contributors built 267 skills in three months, 31% of R&D contributing actively. Adoption spilled past engineering, with 1,100 «Claude Code» users among 1,305 employees.

Curran's own method is anti-perfectionist: don't search for perfect measures, embrace imperfect ones, monitor holistic outcomes. The «Ramp» convergence sits alongside this as a second data point from a scale-up publishing its receipts. What remains open is whether AI-approved PRs actually cross 50%, which Curran predicts but has not yet shown.