Linear Says AI Now Writes Almost Half of All Issues, Yet Teams Ship Slower

[WCWAlfie14]

Linear's own data shows that AI now authors just under half of everything created inside the project management tool, up from less than one issue in a thousand just two years ago. At the current growth curve it is expected to soon author more issues than people and integrations combined, which is a genuinely wild trajectory for something that barely existed as a category before.

At the same time, separate benchmarking from LinearB analyzing over eight million pull requests across thousands of organizations found that developers using AI complete more tasks and merge far more pull requests, yet actual measured delivery time has gone up rather than down. The gap between how fast developers feel and how fast they actually are was described as a thirty nine point spread between perceived and real productivity.

The mechanism behind that gap seems pretty clear once you look at the size numbers. AI assisted pull requests run about two and a half times larger than purely human written ones, and larger diffs mean more cognitive load for whoever has to review them, more surface area for subtle bugs, and slower approval cycles even when the code itself was generated instantly.

What is interesting is that Linear frames this less as an execution problem and more as evidence that AI has changed how teams build far more than how they decide what to build. Planning time per user held basically steady even as everything else moved, which suggests the actual bottleneck in software has quietly shifted from writing code to reviewing and deciding what to do with the flood of code AI can now produce on demand.

The honest takeaway from this data seems to be that generation speed and delivery speed are no longer the same thing at all, and treating them as interchangeable is probably where a lot of the current AI productivity hype falls apart in practice

How teams build - Linearlinear.app

[Quanta]

The perceived versus actual productivity gap is the single most important stat in the entire AI coding conversation right now and it barely gets discussed outside of niche engineering circles. Everyone feels faster because typing feels faster, but feelings are not the same thing as shipped features actually reaching users.

[Cached Stephen]

Two and a half times larger pull requests explains basically everything downstream here. Bigger diffs are inherently harder to review properly no matter who or what wrote them, and it seems obvious in hindsight that this alone would eat any raw speed gains from generation.

[CosmicRay67]

My team went all in on AI generated PRs about six months ago and I can personally confirm the review bottleneck is brutal now.
We used to trust a clean looking diff pretty quickly, but AI code that looks confident and idiomatic on the surface sometimes hides a genuinely wrong assumption buried three functions deep that takes way longer to catch than an obvious human mistake would.

[Florence19]

Almost half of all issues being AI authored within two years from basically zero is one of the fastest technology adoption curves I have ever seen tracked in any industry, software or otherwise. Feels like it deserves way more attention than it is getting outside of dev circles specifically.

[Cobra]

The bit about AI PRs waiting longer before anyone even starts reviewing them but then getting reviewed faster once someone finally does is such a weirdly specific bottleneck. Feels like a trust issue as much as a workload issue honestly, reviewers procrastinating on AI code specifically because they know it will take real mental effort

[Megan95]

The planning time staying flat detail is fascinating to me because it implies AI has not actually made teams any better at deciding what matters, it has only made them faster at executing on whatever they already decided to build. Execution speed without better judgment upstream just means you ship the wrong thing faster than before.

[Ethan93]

Worth remembering Linear has an obvious incentive to make their own AI agent usage numbers look impressive since they are actively building agentic features into their product. Doesn't mean the data is fake, but it is worth reading the framing with a little skepticism about who benefits from the narrative.

[Chris81]

This tracks with something a senior engineer told me a while back, that junior mistakes look like junior mistakes and are easy to spot, but an AI's mistake looks like confident senior level work wrapped around a hidden flaw. That asymmetry is genuinely underrated as a risk in all of this hype

[MondayMoan]

Curious how much of this slower shipping despite more output is actually a temporary transition cost rather than a permanent state. Teams are still figuring out review processes built for human paced code, and it would not surprise me if this gap closes somewhat once better tooling for reviewing large AI diffs specifically starts to mature