How Intercom Builds with AI, with CTO Darragh Curran

Intercom is a customer support software used by orgs like Microsoft, Anthropic, Perplexity, Vanta, and Clay. Every month, they serve over 600M end-users! The Intercom team has spent a lot of time thinking deeply about AI’s impact on customer support. They were one of the first movers in the space when they launched Fin: their AI customer service agent, earlier this year. 

We spoke to Intercom’s CTO Darragh Curran to find out how his team actually uses AI. 

Here are the highlights!

AI as a product, not a plugin

Intercom has been shipping AI features for years, but the step change with modern LLMs unlocked a completely new set of possibilities. Instead of bolting AI onto workflows, they built the infrastructure to test, measure, and iterate - hundreds of A/B tests, all compounding over time.

That’s how Fin’s resolution rate grew from ~26% to over 50%, often improving a percentage point each month. The process goes beyond just swapping in new models. It focussed on treating AI like a product surface to conduct experiments. 

Measuring engineering velocity 

We asked Darragh if AI has made their engineering team faster. Darragh was candid here: he said coding feels easier, but bottlenecks have moved elsewhere. Specifically decision-making, wait times, review loops. So they measure what matters:

• PR throughput per engineer.
• Feedback wait times (tests, reviews, deploys). If someone waits six hours for a review while another waits 30 seconds, they investigate.
• % of code changes co-authored by AI. Today, ~1–2% of changes are fully written by AI; a much larger share is AI-assisted. 

In short, AI can help increase velocity only if you know where the bottlenecks really are.

Intercom’s internal dev tools team 

Inside Intercom, there’s a team dedicated to making AI the default developer environment. They run autonomous jobs that remove dead code, clean up stale flags, and file PRs: tedious but important jobs no human wants to prioritize.

Over time, these automations evolve into repo-wide refactors and eventually bigger lifts like framework swaps. Think of it as an AI dev-tools startup inside the company, proving safety with narrow scopes before expanding to higher-stakes jobs.

“There’s valuable work AI can do that humans will never prioritize - even though it compounds.”

Creating a shared repository of AI engineering best-practices 

Some engineers intuitively use AI tools; others need a practical framework to upskill. Darragh expects the team average to rise a full point on their internal “impact scale” just by evenly distributing know-how, even if the tools stay static. 

They proactively standardize prompts, document repeatable wins, and review diffs as a team so these patterns spread.

“If you’re not using these tools, you’re probably underperforming - and you’ll get left behind.”

Choosing metrics that matter

Intercom rejects “feel-good” dashboards. They use behavior-shaping metrics:

• Resolution rate (for agents like Fin), constrained by accuracy.
• PR throughput, review latency, deploy cadence (for engineering).
• AI contribution share (to set expectations and track adoption).

Set thresholds that trigger action (e.g., review waits > 2 hours = auto-escalation). Make them visible so teams self-correct.

He thinks teams should publish a weekly metrics memo that documents resolution rate changes, top latency offenders by team, percentage of AI-assisted changes, and experiments that moved the needle.

Accuracy over spectacle

Intercom’s biggest lesson with Fin was: correctness beats cleverness. Customers don’t care if an agent sounds smart. They just want to have their problem resolved quickly. Yes, the magical moments AI can create for the customer matter, but every UX win at Intercom is paired with a truth metric: hallucination rate, retrieval coverage, source accuracy.

When a change boosted deflection but hurt accuracy, they rolled it back. A customer’s trust is harder to earn than it is to lose.

Takeaways for builders

• Build the engine, not the demo. An A/B infrastructure, guardrail metrics, and repo-wide automation paths matter more than the flashy first win. 
• Fix the queues. If AI doesn’t move velocity, your bottleneck is in reviews, tests, deploys, or prioritization.
• Automate the grunt-work. Start with dead code and flag cleanup; graduate to refactors. The compounding ROI is real.
• Codify usage. Make AI-first environments default and share patterns until the average rises.

These principles have led to steady, compounding improvements in both customer trust and developer productivity.

We had a great time jamming with Darragh! We’ve been having similar conversations with product and engineering leaders at top engineering orgs (like Vercel, Wix, Vanta, etc) to unpack how they actually build with AI. You can find previous episodes on YouTube.