Why AI-assisted PRs merge at half the rate of human code | LinearB’s 2026 Benchmarks

0:04Ben Lloyd Pearson:Hey, what's up, everyone? Welcome back to Dev Interrupted. I'm your host, Dan Lines, LinearBee, COO and co-founder. Today's episode will focus on the 2026 Engineering Benchmark Report from LinearBee. So this year's report, it's our biggest, our most comprehensive yet. It's an amazing, amazing report. we're going to take a look at what the data says about engineering teams and really specifically how AI is fundamentally reshaping the way we build software.

0:41Dan Lines:And to help me walk through all this

0:44Ben Lloyd Pearson:data, I have my fellow co-host and Linear B's Director of AI Innovation, Ben Lloyd Pearson, Jason BLP to answer all of my questions about this year's report. Ben, always great to share an episode with you. Welcome to the show.

1:03Dan Lines:Yeah, thanks for having me. It's always a lot of fun to just share all the cool research we're doing at Linear B, so a lot of cool data for us to unpack today.

1:12Ben Lloyd Pearson:Yeah, amazing. And you and I were talking before the show. You came on, Devinterrupted, to do this, I think, like two years ago, maybe even three years ago. Three years ago. Three years ago. And so I think it will be pretty interesting to see how this has progressed. Obviously, a lot of it has to do with AI. But how do you feel about now coming on three years later and doing the same thing, but with an updated report?

1:41Dan Lines:Oh, man. I mean, it's wild to think about just how different the world is today versus when I first started here at Linear B. I mean, And AI was barely even something we were talking about at that point. And now it's like all we can talk about. And I feel like everything I'm doing is now being impacted by it. So, you know, of course, so is this report. And there's going to be a lot of new AI stuff that we're going to talk about today, which is pretty awesome.

2:05Ben Lloyd Pearson:Yeah, it's really cool. I mean, I read through everything in there and definitely like it's going to be exciting to get into the AI stuff. That's really popping out. A lot of good insights there. But I guess probably the best way. Let's start with an overview. Let's start out with an overview summary. Maybe you can give us an overview of what makes either this report unique or maybe some of the things that are kind of like jumping out at you for this year.

2:33Dan Lines:Yeah, well, like you said, it's the most comprehensive analysis we've ever done. 8.1 million poll requests, about 4 ,800 engineering teams spread across 42 different countries. So pretty large scale of data that we're dealing with here. And what really sets this year apart is that we've introduced a completely new dimension to this report around AI productivity and specifically how to measure it. So for the first time, we're not just looking at all the traditional software delivery metrics. You know, those are still there. We're still reporting on them. But we're also examining how AI is impacting engineering workflows across the board.

3:11Dan Lines:So, and on top of that, we've combined for the first time ever qualitative data along with our, excuse me, qualitative analysis along with our quantitative data, where we surveyed a bunch of engineering leadership or engineering leaders about how they're using AI within their organization, both to understand the data behind what's changing, but then also the perception of the leaders that are running these things. And if I had to pick one thing that is sort of the top line that everyone should take away from this is that AI adoption is effectively maximized at this point. It's almost universal.

3:47Dan Lines:So in our findings, 88.3 % of developers now use AI regularly. So that's, you know, at least multiple times a week, which is up from 70, just under 72 % when we last surveyed this back in early 2024. But I want to stress one really critical point, and that is that adoption does not equal impact. So AI-generated PRs are behaving completely different than ones that are unassisted by AI. They're larger, they wait longer for reviews, and they merge at less than half the rate of human-authored code. So what I really want people to take away is that AI is accelerating code generation. I think we all know that at this point, but it's also exposing bottlenecks everywhere else in the SDLC, primarily with things like reviews, testing, governance, organizational readiness.

4:41Dan Lines:There's still a lot that needs to be solved outside of code generation.

4:46Ben Lloyd Pearson:Yeah, man, you mentioned a few things there. One thing that I really like about the report, and hopefully like all of our listeners will read through this, we do have the qualitative responses now. So as you read through the report, you can see what engineering leaders are saying and commenting on and that type of thing. So I thought that was really cool. The other thing that I really liked about the report on the AI side is it doesn't just talk about AI, I guess, generically as being like one single unit. it's there's the agentic pull requests there's the ai assisted and then there's the non ai assisted

5:27Dan Lines:and the comparison there exactly um maybe you can explain you know what what's the difference

5:33Ben Lloyd Pearson:between like the ai assisted and the fully agentic just uh for our listeners yeah yeah so we have a

5:39Dan Lines:lot of data that compares these three classifications of prs so um you know on the simplest side you have unassisted PR. So this is something where a human authors it, they don't use AI to generate code at all, probably don't even use it to ideate, they just write the code sort of as we've always done for many years, and then submit it. One level up from that is AI assisted PRs. So this is where the code is authored by a human, but it's significantly shaped with AI tools, whether that be the actual generation of the code, or, you know, maybe planning for generating the code, or research that type of stuff.

6:16Dan Lines:And then at the highest level, you have agentic PRs. So those are pull requests that are created entirely by an AI agent. Definitely the least mature of these three categories. But that's agents like Devin, Copilot, stuff like that, when they generate the pull requests themselves rather than having a human lead the effort.

6:36Ben Lloyd Pearson:Yeah. And we're going to get into some of those differences because I thought that was really cool how the report broke that out. And obviously that is new to the report. What else is new compared to previous years for the report?

6:52Dan Lines:Yeah, so the biggest thing is probably the new section that we introduced that focuses exclusively on AI productivity insights. And we've broken this down into a few parts. So AI code in the SDLC. So this is how AI generated and AI assisted PRs behave compared to manual or unassisted pull requests. the devx of AI. So this is how developers are experiencing AI adoption and their level of trust with it. And then the last big addition related to AI is a new survey that gauges the state of AI readiness among respondents. So this is a look at whether organizations have the foundations to make AI successful.

7:34Dan Lines:And if you've read this year's Dora report, this will be very familiar with you because we tried to build upon the great research that they're doing over there. Related to all this new AI stuff, we've also added a new benchmark this year for acceptance rate. And that's not the typical one of how many lines of code did you accept or what percentage of recommendations from a coding assistant did you accept. It's not that. Instead, we're measuring the percentage of pull requests that get merged within 30 days. So this is really relevant for AI generated code where things like ownership and acceptance patterns can be completely different than your typical pull request.

8:16Dan Lines:And, you know, just a preview before we get into it too much, the data on this is really striking. You know, manual PRs merge at about 84.5%. So about 84 % of all the unassisted PRs that are open get merged across all of our users. But AI PRs merge at just 32.7%, which is less than half of that rate. So a pretty stark difference between those two numbers.

8:45Ben Lloyd Pearson:Yeah, and it's like super cool that we have that data. That's some of the ones that popped out to me. And yeah, when we say acceptance rate, usually that's something that's pushed by the vendors, like Copilot and all of them saying how much, you know, code that we suggested to developers that they actually accept. I think this is even actually more useful. It's more so saying of all this code that's being generated, how much of it is actually making it into customers' hands or like into production. And there's such a big difference between the agentic, the assisted, you know, fully developer created.

9:22Ben Lloyd Pearson:I know we're going to dive into it, but I really wanted to highlight that one as well. Yeah.

9:27Dan Lines:And I mean, when you're thinking about lines of code that is generated by an AI, if you measure that as your acceptance rate, it doesn't actually mean that any of that code ultimately gets accepted. The developer might just be generating it and then deleting half of it because they don't like it. Our version of this metric is a great way to understand of all the things that AI is generating for your organization, which of them are actually useful enough to be deployed to production. That's a far more meaningful metric to track.

9:54Ben Lloyd Pearson:Yeah, high value metric. Okay, I know there's a few other things that are different, or maybe different in how we're analyzing the data. We already talked about, okay, now there's three types of PRs, right? There's the agentic AI PRs, there's the AI assisted PRs, there's the unassisted, so like fully developer PRs. I think there's something about an AI readiness matrix. Yeah, maybe talk to us about that and if there's anything else that you want to dive into.

10:27Dan Lines:Yeah, I'll cover the readiness matrix, but I do want to point out one, just one really interesting factor about those three classifications of PRs that we described. So again, that's unassisted, that is AI assisted and agentic AI PRs. You know, like I said, we wanted to track differences in behaviors between all of these. And some examples of what we were able to discover is that, for example, with AI-assisted PRs, they tend to be about two and a half times larger than unassisted PRs when you're looking at the P75, so the 75th percentile. But what's more striking is that AI PRs have a pickup time that is more than five times longer than unassisted PRs.

11:12Dan Lines:So they're basically waiting idle for review much, much longer than manually generated code. So again, very stark differences in what we're seeing here. And then on top of that, we've also introduced this AI readiness matrix. Again, based on Dora research, within the latest Dora report, they have these seven categories of success criteria with AI. A big story they're pushing with Dora is how AI amplifies both the good and the bad. So if you have, for example, bad version control practices, when you start using AI, it's going to amplify all those bad sides of it. So we surveyed a bunch of engineering leaders to understand whether the typical organization has the foundational capabilities.

11:58Dan Lines:So there's things like data quality, version control, maturity, clear AI policies, the things that actually takes to make AI successful. And there's some pretty eye-opening findings around that too. For example, 65 % of companies lack dependable data quality. And on top of that, AI policy alignment is sharply polarized. So there is a very big difference between organizations that are strong at this stuff and not so strong.

12:28Ben Lloyd Pearson:Okay. Amazing stuff. New report. A lot of new things going on in the world with AI. The report captures all of the differences. the first one that I'd like to dive into is what you were giving us, I think, a little appetizer on with the unassisted versus the assisted versus fully agentic. Maybe you can recap again the data behind that. But what I want to get into is let's figure out why these, because the behaviors are like so different there. So maybe give us a recap of the data and then let's talk about like why that's happening.

13:04Dan Lines:Yeah. So to recap, AI-assisted PRs are being accepted and merged at about half the rate of manually generated PRs. And we had a couple of other data points that sort of indicate why this might be happening. So again, looking at the P75 level, the 75th percentile, the typical AI assisted PR has a little over 400 lines of code, which is pretty large. It's not gigantic, but it is large. And that's compared to 157 lines of code for unassisted PRs. And then agentic AI PRs are kind of in the middle at about 290-ish lines of code. So when you think about it, one of our recommendations we've always had at Linear B is that you want to try to keep your PRs below 300 lines of code because that's sort of a chunk of information that is relatively, it's manageable for a human to keep all that information in our head and make clear decisions about it and review it promptly and efficiently.

14:03Dan Lines:So when you have pull requests that are getting, you know, 400 plus lines of code, it increases a lot of the mental tax on people who have to review the code. You know, so people are reviewing more, there's more files they have to look at, there's more parts of the system that get affected when PRs get bigger, and just an overall higher complexity. And I really love the qualitative survey we introduced with this report because it does add a lot of flavor to all of the data points that we've been gathering to help us understand why these sorts of things are happening. And the survey of questions around this kept coming back to very similar issues.

14:42Dan Lines:Like AI tends to make changes that are larger than the scope of what was requested from them. They're very verbose at times. They often just do more than a human would, which is, I mean, when you think about it, creating more code is effectively free for an AI. It takes a lot of mental power for a human to do it. So it's just naturally much easier for an AI to just generate larger volumes.

15:05Ben Lloyd Pearson:Yeah, I mean, I think the takeaway is like, OK, if I'm a developer, I'm not using AI. My PR sizes are generally smaller, probably because, you know, it's pretty intensive to actually code. You got to write, think about it, write out all the code. As a developer, you're usually trying to get the most done possible with the least amount of code. I mean, that's really hard hands-on work. For AI, these PRs, in one sense, could be getting bloated. And the downstream impact of that, I'm thinking about quality. I'm thinking about quality both from, I don't know, maybe bugs and security and that kind of stuff.

15:43Ben Lloyd Pearson:But I'm also interested in the reviewer portion, meaning if a human's going to go and have to review that. Now, maybe AI is also doing the review, but if a developer is going to review it, well, my review is definitely not going to be as good as if it's smaller. And then you also mentioned something about maybe some scope creep, too, in there. Like, is it doing more than it's really asked to do? It feels like it's putting so much pressure on the review process. Is that how you feel about it, or what are your thoughts?

16:16Dan Lines:Yeah, I mean, that is the bottleneck right now that I think most organizations that are adopting AI are struggling with the most right now. So, you know, I already mentioned that these that AI generated PRs wait over five times longer. It's actually about 5.25 times longer to be picked up for a review. So that's a representation of a little over a thousand minutes, which a smarter person than me would have to figure out how many hours that is for AI generated PRs versus about 200 minutes for an unassistant PR to get picked up for review. But here's where things get really kind of weird. It will make sense, I think, if we break this down.

16:54Dan Lines:Once someone starts reviewing an AI-generated code, it tends to get reviewed much faster. In fact, the typical AI-assisted PR gets reviewed in about 194 minutes compared to 252 minutes for manual PR. So while they take longer to get picked up, they get reviewed faster.

17:18Ben Lloyd Pearson:Okay, yeah, we got to break. That's really interesting. so the pickup time is longer meaning okay let's say I see that there's a PR out there and it needs to be reviewed if it's assisted by AI it takes longer for that review to actually begin but once the review begins it's actually faster do you have a guess why well I mean I would say that I'll start with the part that makes most sense to me because right now we're talking about assisted versus unassisted but there's that third category of full of fully agentic right that third category of fully agentic i could see why those wouldn't be picked up as fast because it's like who who owns that or like uh i don't feel like blp if you opened up a PR that you worked hard on and then it was like assigned to me and we're buddies and I know you're trying to get this done.

18:25Ben Lloyd Pearson:I have a lot of incentive to go help out my boy like to, okay, I got to get BLP's code out there. That makes sense to me because there's like a human component in ownership. Now, if AI, which I have no like a relation to, opens up to me what's a random PR, like I'm not really incentivized to pick it up. Like that makes sense to me. Yep. The only thing that I could say, now, let's say, BLP, you're using assistance in your, so let's say that you're using Copilot, but it would still be under your name when the PR goes up. I would still be incentivized to go get it, but we also said that the PR is larger.

19:09Ben Lloyd Pearson:And if the PR is larger in general, me as a reviewer, I'm going to be less incentivized to want to start that. That's the only thing I can think of. Is it that or is there other stuff behind the scenes?

19:21Dan Lines:You're definitely on to something. And I think there's a little more to this too. So, and again, this is where it was really great for us to lean on qualitative data this time as well, because one thing that came up in multiple survey respondents was that, you know, a lot of people are just hesitant to open AI PRs, you know, they're uncertain about them. They don't know what the mental load is going to be on that. They have concerns about trust with the AI, like is it going to contain errors or missing context? And then it becomes my problem. Like if I look at it and I start reviewing it, then it becomes my problem, you know.

19:54Dan Lines:So I think that explains part of why they take so long to get picked up for review. But then once the reviewer does pick up, you know, they tend to just scan for high-level issues rather than like deeply understanding the change. You know, one of the respondents to our survey said that, you know, a larger amount of our code is slipping into production without proper review. So rubber stamping, like that's what is starting to happen more and more. And then we also heard concerns about how AI often gives up mistaken suggestions and non-working code. And I've been hearing a lot, particularly in the last year, about the struggles with trust when it comes to AI performance.