AI Changed How Spotify Builds: What We Learned and Fixed About Quality at Higher Velocity

Spotify is now moving faster than its quality controls in a few different places at once, and that’s the real story here. The company says AI has raised the pace of change across development, but the problems it’s dealing with are not some dramatic new plague of AI slop. They’re the old, familiar kind: missed alerts, capacity crunches, and checks that lag behind the speed of shipping.

The most concrete example came in content processing. Spotify handles more than 500,000 new songs, videos, podcasts, and audiobooks every day, and on June 24 a set of small issues combined to delay episodes for hours. A scheduled batch job collided with new uploads, a recent improvement had made each episode more expensive to process, and a scheduler bug cut throughput by about 10%. Spotify says it has since added end-to-end monitoring, fixed the scheduler, lowered the priority of batch jobs, increased capacity, and changed service tiering so new uploads and critical services win when capacity gets tight.

Automation brought its own problems. Spotify’s Fleet Management system has long pushed large-scale changes through with most of them merging automatically after safety checks. It has now been extended to more complex agentic changes, including a Java migration across backend services that finished in three days. But an automated dependency upgrade still passed checks and failed in production this year, so the company is tightening safeguards, expanding rollback capacity, and scheduling automated changes during the owning teams’ working hours.

Then there’s compute. AI demand across the industry has eaten into spare CPU and GPU capacity, and Spotify says that made regional failovers more painful earlier this year. Events that used to be trivial became visible to users because the backup region didn’t always have room for lower-tier services. The company doubled reserved edge capacity after a May incident and is still extending traffic controls and testing spillover.

Spotify’s own data points to a subtler picture than the usual AI panic. In the incidents it reviewed, it did not find AI-authored code as a material direct cause. What it did find was that the volume of change rose faster than review, testing, rollout, observability, and rollback could absorb. Merged changes more than doubled year over year in August, from roughly 8,100 to 17,000, and quality and optimization work rose from 27% to 31% of the mix. Meanwhile, rework rate did not show the kind of rise Spotify would expect if it were accumulating AI-induced quality debt.

The warnings are less dramatic but more interesting: code complexity and PR size are creeping up, and Spotify isn’t rushing to declare that a problem or a victory. That restraint is the right move. AI is making teams faster; the boring, expensive part is proving the systems around them can still tell a good change from a bad one.