How Building with AI Can Double the Throughput of Your Engineering Team — Brian Scanlan, Intercom
TL;DR
Intercom set an explicit goal to double engineering throughput in one year — and says it hit 2x PR throughput in less than 12 months. Brian Scanlan says the company tracked “code changes per R&D person” as the main metric and saw the inflection after standardizing on Claude Code in December and rolling it out in January.
The big move wasn’t just buying copilots — it was treating AI adoption as an org change program with hard expectations. Intercom updated job descriptions so that not adopting AI means “not meeting expectations,” staffed a full-time 2x team, and reinforced the push through hackathons, immersion days, and constant executive repetition.
Intercom picked one agent platform on purpose, even though employees were already using Cursor, Copilot, and Augment. Scanlan compares multi-tool AI sprawl to multi-cloud: you miss the compounding benefits of a shared platform, so they went all-in on Claude Code and optimized it deeply for Intercom’s environment.
Their core idea is to make Claude act like a senior engineer across the whole company, not just autocomplete code. That meant connecting it to everything an engineer can access, teaching it Intercom-specific Rails conventions, security rules, and architecture, and building internal skills, hooks, and plugins so it can debug, plan, test, and review code.
The most memorable example was a security incident Claude handled in about 2 minutes instead of 20. Scanlan dropped the agent into a Slack channel about accidentally exposed Snowflake table metadata, and it automatically invoked an internal data-breach skill, analyzed the files, judged the case innocuous, and laid out next steps without being explicitly told how.
Intercom says AI is now reshaping bottlenecks, quality, and compliance — not just speed. About 17.6% of PRs are auto-approved through heavily tested review agents, the company says this still satisfies SOC 2, ISO 27001, and HIPAA requirements, defect backlogs are getting cleared faster, and Stanford researchers reportedly found code quality rising on their metrics.
Summary
Intercom’s AI reinvention, in numbers
Scanlan opens by positioning Intercom as a 15-year-old Irish-American B2B SaaS company that “pivoted to be an AI company the weekend that ChatGPT came out.” He runs through the receipts fast: 1,400 people across Dublin, London, Berlin, San Francisco, Chicago, and Sydney; AI support agent Finn with 8,000+ customers; revenue approaching $100 million; and roughly 2 million resolutions a week.
Shipping is the heartbeat — so productivity became the mission
As a senior principal engineer on Intercom’s platform group, Scanlan frames the company’s culture around shipping quickly and iteratively. By mid-last year, the team felt tools like GitHub Copilot, Cursor, and Augment were helpful but still too marginal, so they set a blunt goal: double engineering throughput in a year, measured primarily by code changes per R&D person.
The leadership playbook: be binary, be repetitive, staff it properly
This wasn’t pitched as a side experiment. Intercom updated job descriptions so that failing to adopt AI means you’re “not meeting expectations,” and leadership kept repeating the message “over and over and over” across every forum. They also celebrated wins publicly, ran hackathons and AI immersion days, and put a growing full-time “2x” team behind the work instead of just telling hundreds of engineers to figure it out alone.
Why Intercom standardized on Claude Code
Scanlan says platform choice mattered less than platform commitment. Intercom chose Claude Code because, in his view, spreading work across too many agents is like going multi-cloud too early: you lose the compounding benefits of a single optimized system. Their ambition was to make Claude behave like a senior engineer at Intercom, with access to the same tools, guardrails, context, and environment as a trusted human developer.
Teaching the agent to think like an Intercom engineer
A huge part of the work was encoding company-specific knowledge: Rails conventions, React patterns, testing standards, architecture, and security rules built up over 15 years. Scanlan says they packaged that context into skills, guidance, hooks, and internal plugins, then pushed those tools aggressively to employee laptops — even bypassing normal update flows because managing installs at scale was “like trying to install Python.”
“Give problems to agents, not tasks”
One of his strongest principles is to stop micromanaging the model. Instead of saying “run this skill,” he wants people to describe the problem and let the agent choose what capabilities to invoke. His favorite example: during a security incident involving Snowflake table metadata in a public GitHub repo, he simply asked Claude Code to join the Slack channel and investigate; it found and used the right internal breach-analysis skill on its own, finished in two minutes, and saved him the usual 20 minutes of policy-checking drudgery.
The inflection point: throughput doubled, and review became the bottleneck
After deciding to go all-in in December and rolling out in January, Intercom saw what Scanlan calls “wild inflection points.” The company claims it reached 2x PR throughput in under a year, and now the limiting factor is increasingly code review — which they’re also automating, with 17.6% of PRs already auto-approved through carefully backtested workflows and agent reviewers.
Quality, defects, and the strange new shape of engineering work
Scanlan argues this shift is improving more than speed. Defects are getting closed faster, some teams are pushing toward “backlog zero,” and Stanford researchers working with Intercom have reportedly seen code quality rise on their metrics. He ends on a broader note: Claude Code has spread beyond engineering, the company’s CI briefly melted under the load, and even single-person product experiments — where one person acts as PM, designer, and builder with agents — are starting to feel real.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Retirement Email Isn't a Warning
Model retirements now arrive every few weeks; the config-eval-rehearsal loop turns each deprecation email from a fire drill into an afternoon swap.
Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.