How Building with AI Can Double the Throughput of Your Engineering Team — Brian Scanlan, Intercom
Summary
Brian Scanlan, Senior Principal Engineer at Intercom, presents the company's journey to doubling engineering throughput in under a year using Claude Code as a unified agentic development platform. Intercom is a 15-year-old B2B SaaS company with roughly 1,400 employees, a Rails monolith at its core, and a flagship AI customer support agent called Finn serving over 8,000 customers at nearly $100M in revenue. The talk centers on the '2x' initiative: a deliberate project to double code output per R&D person without doubling headcount.
Scanlan opens by contextualizing Intercom's AI-first transformation. Unlike companies adding surface-level AI features, Intercom pivoted hard the weekend ChatGPT launched, shipping the first product on GPT-4 on launch day. The company now runs a proprietary model serving 100% of Finn's English text conversations, outperforming frontier models at lower cost, processing roughly 2 million resolutions per week.
On the internal engineering side, the 2x project started mid-2024 when Scanlan's team felt dissatisfied with GitHub Copilot, Cursor, and Augment — seeing only marginal gains. The decisive shift came in December 2024 when they chose Claude Code as the single platform for all engineers, rolling it out in January 2025. Scanlan emphasizes the 'platform commitment' principle: going all-in on one tool compounds benefits rather than fragmenting attention across multiple agents, analogous to choosing a single cloud provider.
Key results reported: PR throughput doubled in less than a year, 17.6% of pull requests are now auto-approved with full SOC 2, ISO 27001, and HIPAA compliance, and Codex is used for multimodal code review within the approval pipeline. The CI infrastructure 'melted' under the surge in volume — an unintended but telling signal of success. Defect closure rates accelerated, with some teams pursuing backlog-zero initiatives. Stanford researchers auditing Intercom's codebase confirm code quality metrics are rising.
Scanlan's onboarding framework for Claude Code mirrors new-hire onboarding: teaching it 15 years of Rails conventions, React patterns, testing standards, and security rules via Claude.md files and skills. The team maintains hundreds of internal skills (encapsulated reusable agent instructions), pushes them to all engineer laptops via internal Claude plugins, and mines session transcripts stored in S3 to evaluate skill effectiveness and guide continuous improvement. A notable principle is 'give agents problems, not tasks' — describing intent rather than prescribing steps. Scanlan illustrates this with a security incident involving accidentally published Snowflake table metadata on GitHub: he described the situation to Claude, which autonomously discovered a relevant data-breach-analysis skill, downloaded the files, ran the analysis, and returned next steps in two minutes — a task he estimates would have taken 20 minutes manually.
The engineering maturity model Scanlan outlines for individual engineers progresses from basic Claude Code usage, through automating personal workflows, to writing and refining skills, and finally to optimizing the entire development environment for agent effectiveness. The team also explores 'single-person team' product experiments and uses Claude Code virally across non-engineering functions. Scanlan closes with a broader career framing: just as Unix sysadmins became SREs when cloud arrived, engineers are speed-running a similar transition — moving up the stack to higher-leverage, higher-impact work.
Key Takeaways
- •Commit to a single agentic platform (Intercom chose Claude Code) rather than letting engineers use multiple tools — platform focus compounds benefits and avoids fragmentation.
- •Treat Claude Code like a new hire: invest in proper onboarding by encoding 15 years of conventions, architecture patterns, security rules, and testing standards into skills and CLAUDE.md files.
- •Frame work as problems for agents, not step-by-step tasks — describing intent lets the agent discover the right skill autonomously, as demonstrated by the 2-minute Snowflake security incident resolution.
- •Auto-approval of pull requests is achievable at scale with regulatory compliance (SOC 2, HIPAA, ISO 27001) if you use rigorous backtesting, human labeling, and confidence calibration before going live.
- •Mine session transcripts (Intercom stores them in S3) to evaluate which skills are actually effective and to drive continuous improvement of the skill library.
- •Staff AI productivity initiatives with dedicated full-time teams — Intercom's '2x team' kept growing because the work is too important to treat as a side project.
- •Use PR throughput (code changes per R&D person) as the primary productivity metric; expect a large, measurable increase — not just marginal improvements — when AI adoption is real.
- •Build small, high-quality, durable, testable skills rather than large monolithic orchestrators; use backtesting against historical work to validate quality.
- •Drive adoption with clear executive mandates (non-adoption = not meeting expectations), visible celebration in Slack, hackathons, and AI immersion days.
- •Expect infrastructure to become the bottleneck — Intercom's CI collapsed under the volume surge, and code review became the next constraint after throughput doubled.
Transcript
[music] >> Hey, I'm Brian from Intercom. I'm this has been such a great conference so far. I've learned so much alpha and inspiration from the talks and all the chats and people. So Intercom is a 15-year-old privately held Irish-American B2B SaaS startup that has pivoted to be an AI company the weekend that ChatGPT came out. I've got about 1,400 people across Dublin, London, Berlin, SF, Chicago, Sydney. R&D is led from Dublin. Engineering is almost entirely across Europe. Four deployed engineers...
Speakers
Topics
Related Sessions
