Does AI Actually Boost Developer Productivity? (100k Devs Study) - Yegor Denisov-Blanch, Stanford

2,991 words · about 12 min to read

Measuring software engineering productivity at scale

0:15 In January of this year, Mark Zuckerberg said that he was going to replace all of the mid-level engineers at Meta with AI by the end of the year. I think Mark was a bit optimistic and he was probably acting like a good CEO would to inspire a vision and also probably to keep the Facebook stock price up. But what Mark also did was create a lot of trouble for CTOs worldwide.

0:44 Why? Because after Mark said that, every single CEO in the world almost turned to their CTO and said, "Hey, Marcus says he's going to replace all of his developers with AI. Where are we in that journey?" And the answer probably was honestly not very far and we're not sure we're going to do that. And so I personally think hopefully this is not um you know going to going to change, but I don't think AI is going to replace developers entirely. at least at least this year, let alone at at at Meta, right? But um and I do think that AI increases developer productivity, but there's also cases in which it decreases developer productivity. So AI or using AI for coding is not a one-sizefits-all solution, and there are cases in which it shouldn't be used.

1:35 And so for the past three years we've uh been running one of the larger uh studies on software engineering productivity at Stanford and we've done this in a time series and cross-sectional way. So time series meaning that even if a participant joins in 2025 we get access to their git history meaning we can see trends of data across time we can see covid we can see AI we can see all of these trends and and things that happened and then also cross-sectional because we have more than 600 companies participating enterprise midsized and also startups and so this means that we have more than 100,000 software engineers in our data set right now dozens of millions of commits and billions of lines of code.

2:22 And most importantly, most of this data is private repositories. This is important because if you use a public repo to measure someone's productivity, that public repo is not self-contained. Someone could be working on that repo on the weekend or once in a while, right? Whereas if you have a private repo, it's much more self-contained and much easier to measure the productivity of a team, of a company, of an organization.

2:49 So late last year there was a huge um controversial uh thing around ghost engineers. So this came from kind of the s the same research group our research group and here uh Elon Musk was kind enough to retweet us. But what we found is that roughly 10% of software engineers in our data set at the time about 50,000 were what we called ghost engineers. These people collect a paycheck but basically do no work.

3:16 So that was very surprising for some people, very unsurprising for others. And so some of the people in this research team are for example Simon from industry. Uh so he was CTO at a unicorn which he exited and he had a team of about 700 developers and as CTO he was always the last person to know when something was up with his engineering team right and so he thought okay how can I change this?

3:42 Myself, I've been at Stanford since 22 and I focus on what I call datadriven decision-making in software engineering and in a past life I was looking after digital transformation for a large company with thousands of engineers. Part of the team is also professor Kasinski who was at Stanford and his research focuses on human behavior in a

Vendor studies and surveys are unreliable productivity metrics

4:04 digital environment and basically he was the Cambridge Analytica whistleblower back in the day if you recall that. So today we're going to be talking about three things. We're going to start off with the limitations of existing studies that seek to quantify the impact of AI on developer productivity. We're going to showcase our methodology. And lastly, we're going to spend most of the time looking at some of the results. What is the impact on AI on deaf productivity? And how are ways we can slice and dice these results to make them more meaningful?

4:37 And so there's lots of research being done on this topic, but a lot of it is led by vendors who themselves are trying to sell you their own AI coding tools, right? And so there's a bit of a conflict of interest there sometimes. And the biggest three limitations that I see is that a lot of these studies revolve around commits and PRs and tasks. Hey, we completed more commits, more PRs. The time between commits decreased.

5:04 The problem here is that task size varies, right? And so delivering more commits does not necessarily mean more productivity. And in fact, what we found very often is that by using AI, you're introducing new tasks that are bug fixes to the stuff that the AI just coded before. So by that case, like you're kind of spinning your wheels in place, right? So that's kind of funny. Secondly, there's a bunch of studies who say, well, we grabbed a bunch of developers, we split them into two groups and we kind of gave one AI and one of them we didn't. And what usually happens there is that these are kind of green field tasks where they're asked to build something with kind of zero context from scratch. And there of course AI decimates uh the non-AI people, but that's because AI is just really good at green field kind of boilerplate code, right? But actually most of software engineering isn't green field and isn't always boiler boilerplate, right? And so there's usually an existing codebase. There's usually dependencies. So these studies

6:05 can't be like applied too too well to these situations either. And then we also have surveys which we found to be an ineffective predictor of

1 Highlight · Developer self-evaluations correlate poorly with actual productivity

6:14 productivity by doing this small experiment with 43 developers whereby we ask every developer to evaluate themselves relative to the global mean or median in five percentile uh buckets from 0 to 100 and then we compared that to their measured productivity. We'll get into what that means later, but what we found is that asking someone how productive they think they are is almost as good as flipping a coin. There's very little correlation, right? And so we found that people misjudged their productivity by about 30 percentile points.

6:49 Only one in three people actually estimated their productivity within their quartile, one quartile. And I think surveys are great. They're valuable for surfacing, you know, morale and other issues that cannot be derived from metrics. But service shouldn't be used to measure developer productivity much less the impact of AI on developers for productivity cases. You can measure

Automated git analysis uncovers net productivity gains

7:12 it to kind of see how happy they are using AI or whatever I suppose. Um great. So now let's dive into our methodology. So in an ideal world, you would have an engineer who writes code and this code is evaluated by a panel of 10 or 15 experts who separately without knowing what every person is uh answering evaluates that code based on quality maintainability output how long would this take me, how good is it, right? kind of like a bucket of questions and then um what happens is that you aggregate those results and we found two things. The first one is that this panel actually agrees with one another. So it turns out that one engineering expert agrees with the other engineering expert when they're talking about an objective code in front of them. And secondly and probably most importantly is that you can use this to predict reality. And reality is predicted by a panel like this. The problem then is that this is very slow.

8:13 It's not scalable. It's expensive. Um, and so what we did is we built a model that essentially automates this, correlates pretty well, it's fast, it's scalable, and it's affordable. The way it works is it plugs into git and then the model analyzes the source code changes of every commit and quantifies them based on a bunch of these dimensions. And then since every commit has a unique author, a unique sha, a unique time stamp, then you can kind of understand okay the productivity of a team is basically the functionality of the code they delivered across time, not the lines of code, not the whatever commits, but the fun like what that code is doing, right? And so then you can kind of put this in a dashboard and uh overlay it across time and get something similar to this.

9:05 Great. So now let's dive into some of our results. So here in September is when this company implemented AI. This is a team of about 120 developers and they were piloting whether they wanted to use uh you know AI in their kind of regular workflow. And we have here um these bars

2 Highlight · The hidden tax of AI-generated rework

9:25 and every bar is the sum total of the output done in that month using our methodology not lines of code. And we can see that in green it's added functionality. In gray it's removed. In blue is refactoring and in orange is reworked. And so rework versus refactoring. They both alter existing code. But rework alters code that's much more recent. Meaning it's wasteful. Refactoring could be wasteful, could be not wasteful.

9:52 And so from the get-go, you see that by implementing AI, you get a bunch more of rework. What happens is that you feel like you're delivering more code because there's just like more volume of code being written, more commits, more stuff being pushed. But not all of that is actually useful. To be clear, I think there I mean based on this chart and overall there is a productivity boost of about 15 to 20%. But then a lot of the gains you're seeing are uh basically this kind of rework which is a bit you know misleading.

10:24 So if I could summarize it into one chart with many discrepancies it would be something like this. So with AI coding you generate or you increase your productivity by roughly 30 40%. Right? You're delivering more code. However, you got to go back and kind of fix some of the bugs that code introduced and kind of, you know, fix the the the mess that the AI made, which in turn gives you an average productivity gain across all industries, all sectors, everything of roughly about 15 to 20%.

10:56 There's a lot of new ones here, which

Complexity and language popularity dictate AI effectiveness

10:58 we're going to see in just a second. So here we have two violin charts and they plot the distributions of the gains in productivity from using AI and so kind of like the y-axis is the gains. It starts from minus 20%, take note, and then it goes up. And here we have kind of four pieces of data being shown. In blue is low complexity tasks and in red is high complexity tasks and kind of like your left uh the chart to the left is green field tasks, the chart to the right is brown field tasks.

11:38 So right from the get-go, the first conclusion we have is that sure it seems like AI performs better in coding with simpler tasks. That's good. It's proven by data. That's awesome. The second thing we see is that hey, it sounds like for low complexity green field tasks, there is a much more elongated distribution and a much higher distribution on average. Keep in mind that this is for enterprise settings.

12:04 This doesn't apply for kind of like personal projects or vibe coding something for yourself from scratch. The improvements there would be much bigger. This is kind of for like real world working company settings. And the third thing we see is that if you look at the high complexity tasks, I mean they're lower than the low complexity ones in average in terms of the distribution, but also in some cases they are more likely to decrease an engineer's productivity.

12:32 Now this decrease could be for many things, many reasons, but that's kind of what we see in the data, right? The underlying causes are still not super clear to us. If we translate this to a chart like this, which is a bit more digestible, you have uh in in the bars and the columns kind of like the average or the median gain and then the line represents the interquartile range. So the bottom of the line is the 25th percentile and the top of the line is roughly 75th percentile. And so here it's very clear to see how we have you know more gains from low complexity tasks, less gains from uh high complexity tasks and then brownfield it's harder to leverage AI um to make increases in productivity there compared to green field.

13:21 So if there is maybe a slide that you could show to your leadership team, it could be this one or it could also be this one. So here we have a matrix

3 Highlight · AI productivity gains by task complexity and maturity

13:29 really simplifying things you know reality is a bit more difficult than this but here we have kind of on one axis task complexity low and high on the other one project maturity green field versus brownfield kind of if we see that hey low complexity green field 30 to 40% gains right from AI high complexity but green field more modest gains 10 to 15 brown field and low complexity pretty good 15 to 20%. And most importantly, high complexity brownfield tasks 0 to 10%.

14:02 These are orientative guidelines based on what you see what we see in the data. And I forgot to mention uh this slide

14:08 has a sample size of 136 teams across 27 companies. So pretty representative. Um and then that's going to derive or this chart is derived from that data. Then here we have a similar matrix except at the bottom we have language popularity. So in low we have examples such as Coobo, Haskell, Elixir really kind of obscure obscure stuff and high is things like Python, Java, you know, JavaScript, TypeScript.

14:38 And what we see is that AI doesn't really help even with low complexity uh tasks for low popularity languages. It can help a bit, but it's not terribly useful. And what ends up happening is that people just don't use it because if it's only helpful two times out of five, you're just not going to use it very often. What's funny or interesting is that for low language popularity and complex tasks, AI can actually decrease productivity because it's so bad at coding in Cobalt or Haskell or Elixir that it just makes you slower. Right?

15:06 Granted this isn't very like this happens but it may be five or 10% of the kind of global development work if that right most of the development work is probably somewhere in the langu in the high language popularity kind of part of the chart and here you have gains between 20% for the low complexity and 10 to 15%

Larger codebases and context windows diminish AI utility

15:26 for the high complexity. So now moving into something a bit more theoretical, less empirically proven, but more so kind of like what we're seeing in in in the data, right? This is like an illustrative chart which has kind of productivity gain from AI on the y-axis and a logarithmic scale of the codebase size, right? From 10,000 lines of code to 10 million on the x-axis. And we see that as the codebase size increases, the gains you get from AI decrease sharply, right? And I think most code bases nowadays are kind of somewhere in the depending on on your use case, right? But they're bigger than a thousand lines of code unless you are a YC startup or something that's like kind of spun out a couple months ago, right? And that's because, you know, there's three three reasons for this really. Context window limitations.

16:20 We're going to see in a second how performance decreases even with larger context windows. The signal to noise ratio is kind of confuses the the model if if you will. And then of course larger code bases have more dependencies and more domain specific logic present. And so then borrowing work from this

4 Highlight · Performance degrades sharply as context windows grow

16:41 paper called no lima which shows you on a scale of 0 to 100 how LLMs perform on coding tasks. You see that as context length increases from 1,00 to 32,000 tokens performance decreases. And so we see all these models here. For example, Gemini 1.5 Pro has a context window of 2 million tokens. And you might think, whoa, I can just throw my entire codebase into it and it's going to retrieve and and code perfectly, right?

17:09 And what we see is that even at 32,000 tokens, it's already showing a decrease in performance from 90% to about 50%. Right? So what's going to happen when

17:19 you move from 32 to 64 or 128, right? You're going to see really, really poor performance here. And so in short, AI does increase developer productivity. You should use AI for most cases, but it doesn't increase the productivity of developers all the time and equally. It depends on things like task complexity, codebase maturity, language popularity, codebase size, and also context length.

17:48 Thank you so much for listening. If you'd like to learn more about our research, you can access our research portal which is software engineeringproductivity.stanford.edu. You can also reach me uh by email or LinkedIn. Super happy to talk about this topic at any time. Thank you so much.