Airbnb's AI writes 60% of its code. The other 40% matters more.
Airbnb's CEO says AI generates 60% of new code and resolves 40% of support issues. What that means for a business running a software or support team, and four things to do this week.
Airbnb CEO Brian Chesky said on the Q1 2026 earnings call that AI now generates 60% of all new code written by its engineers. If you run a team that builds software or answers customers, the useful number is the remaining 40%: the work AI has not absorbed, and where your people now earn their pay.
Airbnb's AI writes 60% of new code and resolves 40% of support issues
Airbnb reported revenue up 18% year over year. Its support bot now resolves 40% of issues without escalating to a human, up from around 33% earlier this year.
Chesky described AI as giving leverage on "supervised tasks": an engineer sets direction and agents execute. Airbnb has not published a breakdown, but the framing points at routine execution: boilerplate, integrations, repetitive logic, standard implementations. The extra capacity goes into more tools for API partners and new capabilities, likely including work Airbnb could not have shipped before.
Chesky also claimed a shift from "a team of 20 engineers" to one engineer supervising agents. That is an earnings-call claim, not a measured ratio.
Customer-facing AI is still unsolved, by Airbnb's own account
He added: "I do not think anyone has figured out AI for travel or e-commerce yet." He listed four failure modes in customer-facing AI: too much text output, no direct manipulation, poor comparison tools, and single-player design for what is often a group decision.
Internal gains (code, support triage) are working; the product side is not. Airbnb is an extreme case, a mature tech company with hundreds of engineers and years of AI investment, but smaller companies see the same split at different numbers.
Your bottleneck moves from execution to judgment
Before AI coding tools, shipping more meant hiring more people to write code. For routine work that constraint is gone. What remains is deciding what to build, what to leave out, when to stop, and whether generated code is any good.
An engineer who produces code fast but cannot evaluate it becomes a liability. One who sets direction, reviews AI output critically and makes sound architectural calls is worth more than before. Support works the same way: once the bot takes 40% of tickets, the remaining 60% are the complex, ambiguous or high-stakes cases, so each agent handles harder problems.
What to do this week
- 01If your developers do not use AI coding tools yet, trial them on one routine workstream.
- 02Review your interview loop.Keep testing whether candidates understand code, since nobody can evaluate AI output without it, but weight raw coding speed less and add an exercise where the candidate reviews and corrects generated code.
- 03List one team's recurring tasks and mark each "clear answer" or "needs context, stakes or nuance".Automate the first list; staff and train for the second.
- 04If you run customer-facing AI, check it against Chesky's four failure modes: walls of text, no direct manipulation, weak comparison, no support for group decisions.