--- title: "AI Made Code Cheap, and I Expect More Jobs | Azendo" description: "AI agents open 36% more pull requests, yet teams ship only a little more. Why I expect more software jobs after the dip, and what to do now." url: "https://azendo.co/blog/ai-engineering/ai-more-software-more-jobs/" --- 1. [Home](https://azendo.co/) 2. [Newsroom](https://azendo.co/newsroom/) 3. [AI engineering](https://azendo.co/blog/ai-engineering/) 4. AI made code cheap, and i expect more jobs # AI made code cheap. I believe that means more software jobs. AI agents now write a lot more code, and teams ship only a little more software. I think that gap is temporary. After it I expect more software, more companies and more jobs. By [Mikkel Schmidt](https://azendo.co/about/mikkel-schmidt/) 13 November 2026 ## Two reports and a question I get every week Two reports came out in the same week of November, and I’ve been thinking about both since. On 9 November two Harvard researchers, Fiona Chen and James Stratton, updated [a working paper on 720,542 software workers at 655 companies](https://fion.ac/jellyfish.pdf). A day later [Gartner forecast](https://www.gartner.com/en/newsroom/press-releases/2026-11-10-gartner-forecasts-artificial-intelligence-will-drive-a-net-increase-of-1-7-million-jobs-by-2030) that AI will create more jobs than it removes by 2030. I started my company in Chiang Mai in 2010\. Most weeks someone asks me if AI will make developers redundant, and lately there’s real worry behind the question. My answer is no. I think we’ll end up with more software jobs than we have today. I also think the next couple of years will be hard for some people. And the Harvard paper contains a model that says I’m wrong. I’ll get to it. ## More code, a little more software The Harvard study is the largest look inside software teams I’ve seen. It uses data from Jellyfish, a platform that companies use to follow their engineering work. That’s 445 million events from GitHub, Jira and calendars, recorded between January 2021 and August 2026\. The authors compare companies before and after they adopted AI coding agents. After a company adopts agents, its developers add 44% more lines of code and open 35.7% more pull requests. That’s the story you’ve heard many times. Then the authors count finished work. Resolved Jira issues go up 13.7%, and they call that result borderline significant. Larger projects and features go up 18%. What changed after companies adopted AI coding agents Lines of code written+44% Pull requests opened+35.7% Larger projects finished+18% Issues resolved+13.7% Headcount didn’t move in any way the authors could measure. Their estimate is a drop of 2.6%, with a margin so wide that they can’t tell it from zero. So there’s a lot more code, a little more software, and the same number of people. If you run an engineering team, you’ll know the feeling. The activity numbers look great, and the list of things you shipped looks about the same as last quarter. ## The queue is at the review desk After the agents arrive, the extra code waits for someone to read it. The time from a review request to an approval grows by 56.1%. Code waits 59.9% longer before anyone opens it, and each change gets 38.6% more comments. One engineer told the researchers that only a single senior colleague was allowed to approve changes on one project. That colleague got to them about once a week. Tools haven’t fixed this yet. Of the companies in the study, 79.1% use AI review tools, and people still write nearly all of the reviews. Maybe the tools will get there. Even then somebody has to be responsible for what goes live, and I don’t see companies giving that to an agent. This matches what I see in our teams. Writing code was never the whole job, and the study puts it at 30 to 40% of an engineer’s time. AI moved the pressure to the people who can judge a change. Does it do what it should, and will it break something next month? Those people are your experienced developers, and their days are full. It also explains something I’ve seen in our own work and heard at several industry events this year: in tech, AI is replacing junior roles first. A junior used to write the simpler code. An agent does that now, and somebody senior has to check it. I went through the trials behind all of this in October, in a piece on [whether AI will replace developers](https://azendo.co/blog/ai-engineering/ai-coding-tools-research/). The short version is that AI is fast at first drafts, and a person is still responsible for the result. ## First comes the dip Gartner’s numbers for this year aren’t pretty. For 2026 it estimates 2.2 million jobs lost to AI worldwide and 1.6 million gained. Another 1.3 million positions were never opened, because executives put the hiring on hold. “Deferred jobs are not the same as lost jobs. They simply aren’t created,” says John Lovelock, chief of research at Gartner. If you’re a graduate looking for your first developer job, that doesn’t help you much. So when people tell me AI is taking jobs, I don’t argue. It is, right now, and as far as I can tell mostly at the junior end. Some good people are having a hard year. Then look at where the same forecast ends. By 2030 Gartner expects AI to have created 10.6 million jobs and removed 8.9 million, a net gain of 1.7 million. In the technology sector the gap is much wider, with 1.9 million jobs created against 0.4 million lost. Gartner's forecast: jobs AI creates and removes by 2030 All sectors, jobs created10.6 million All sectors, jobs removed8.9 million Technology sector, jobs created1.9 million Technology sector, jobs removed0.4 million Gartner also expects just over 20% of all workers, 435.5 million people, to see their job change shape. That’s about one in five of us. The job stays, and the work in it changes. ## Where I disagree with the Harvard model A forecast is only a forecast, and it would be cheating to quote the one that agrees with me and stop there. The Harvard authors built a model on top of their data. It asks what happens to a company’s headcount once the company has fully adjusted to AI. The answer depends mostly on one thing, which is how much more software customers buy when it gets cheaper. If they buy a lot more, the company hires. If they buy a little more, it needs fewer people. For that number the authors use earlier research on mobile apps, retail software and console video games. With the average of those estimates, a company that adopts AI ends up with 16.1% more output and 13.2% fewer employees. I take that seriously. They are careful researchers, and I’m a guy who runs a software company. But I think those estimates look at the wrong market. Apps, boxed software and console games are finished products sold to many buyers. I’m talking about software built for one business, and I doubt those numbers say much about that market. The customers I think about never bought custom software at all. Think of a freight company that runs on spreadsheets and phone calls, or a clinic with a booking system that almost fits. For them, a system built for their business used to cost more than the problem did. When AI can bring a build down to a fraction of the time and cost, they become buyers for the first time. The US Bureau of Labor Statistics makes a similar point in drier language. In [an article on its methods](https://www.bls.gov/opub/mlr/2025/article/incorporating-ai-impacts-in-bls-employment-projections.htm) it says AI productivity “may lower prices and increase demand for software products”. It [projects 10% more jobs](https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm) for software developers and testers between 2025 and 2035\. The bureau has lowered that projection three releases in a row, and it’s still growth. I can’t prove my side, so call it a bet. It’s one I’ve put my company behind. ## Everyone can bake bread > Everyone can bake bread. That doesn't mean everyone does. You still buy it, and bakeries still do good business. Software will be the same. Mikkel SchmidtCEO and founder, Azendo AI is smart, and yes, it has opened software up to everyone. A founder with no technical background can now get a working app out of a chat window. It’s a strange time we live in. But being able to build software doesn’t mean you want to spend your days on it. Most businesses will keep their attention on what they sell. They’ll hire someone to build their software, the same way they hire an accountant or buy their bread. What really changes, as I see it, is the price. Custom-built software used to be for companies with big budgets. Now everyone will be able to afford it, and I find that exciting. It has happened before. When cash machines spread through American banks, you’d have expected bank tellers to disappear. James Bessen of Boston University [looked at what happened](https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm). The number of tellers a city branch needed fell from 20 to 13 between 1988 and 2004\. A branch became cheaper to run, so banks opened 43% more of them in cities. Through those years the teller jobs stayed, and the work moved from handling cash to looking after customers. I expect software to go the same way. When a build costs less, more gets built, and all of it has to be reviewed, tested, secured and kept running. Somebody also has to decide what gets built in the first place. How I expect cheaper software to turn into more work 1. CostA build takes a fraction of the timeAI agents write the first draft of each change, and an engineer approves it. 2. BuyersNew customers can afford itBusinesses that ran on spreadsheets and off-the-shelf tools order software built for them. 3. ProductsFar more software goes liveEvery new system has users, data and security to look after. 4. JobsMore people are needed around itSomeone reviews, tests, runs and improves each product, year after year. ## What I’d do with this today If I ran an engineering team at another company, I’d start by measuring finished work. Look at what went live this month, and how many days each change sat in review. The time from a pull request opening to its first review is a good place to start. Lines of code and pull requests only tell you how busy the agents are. Then I’d protect the reviewers. Your most experienced developers are now the bottleneck. Give them time for review, and keep changes small so each one is quick to read. Add people who can review before you add more agents. I’d keep hiring juniors too, fewer than before and with a plan for each one. The seniors who approve your code in 2030 are juniors today. They only get there by doing real work next to someone experienced. Our COO, Kasper, thinks the pressure on junior jobs will even out over the next couple of years. I agree with him. And if you once dropped a software idea because the quote was too high, ask for a new quote. I’d expect a different number. If review is where your team is stuck, this is what we do at Azendo. We put together a [dedicated software development team](https://azendo.co/services/dedicated-software-development-team/) for your product. The people in it are our offshore developers in Chiang Mai and Bangkok. All of them have been through our AI training, and the day-to-day management is on us. Companies that want something built can use our [custom AI software development](https://azendo.co/ai-services/ai-software-development-services/). The method behind it is in our free book, [The AI Software Factory](https://azendo.co/blog/learning/ai-software-factory/). I may have the timing wrong. Gartner will update its forecast and the Harvard authors will update their paper. When they do, I’ll update this post. And if you see it differently, [I’d like to hear why](https://azendo.co/get-in-touch/). ## Sources The figures above come from these publications, as they stood on 13 November 2026. 1. Fiona Chen and James Stratton (Harvard University), [Artificial Intelligence in the Firm: Bottlenecks in Software Production](https://fion.ac/jellyfish.pdf), working paper, version of 9 November 2026. 2. Gartner, [Gartner Forecasts AI Will Drive a Net Increase of 1.7 Million Jobs by 2030](https://www.gartner.com/en/newsroom/press-releases/2026-11-10-gartner-forecasts-artificial-intelligence-will-drive-a-net-increase-of-1-7-million-jobs-by-2030), press release, 10 November 2026. 3. James Bessen, [Toil and Technology](https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm), Finance & Development, International Monetary Fund, March 2015. 4. Machovec, Rieley and Rolen (US Bureau of Labor Statistics), [Incorporating AI impacts in BLS employment projections: occupational case studies](https://www.bls.gov/opub/mlr/2025/article/incorporating-ai-impacts-in-bls-employment-projections.htm), Monthly Labor Review, February 2025. 5. US Bureau of Labor Statistics, [Occupational Outlook Handbook: Software Developers, Quality Assurance Analysts, and Testers](https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm), August 2026. Send us a message Tell us what you want to build. We come back with the team we would assign and what it costs. ## Questions about AI, shipping speed and software jobs. Why don't teams ship much more software when AI writes more code? Because the code still has to be reviewed by a person. In a Harvard study of 655 companies, AI coding agents raised pull requests by 35.7%, while resolved issues rose 13.7%. The time from review request to approval grew by 56.1%. Review is where the extra code waits. How many jobs will AI create and remove by 2030? Gartner's forecast of November 2026 says AI will create 10.6 million jobs and remove 8.9 million worldwide by 2030, a net gain of 1.7 million. In the technology sector it expects 1.9 million jobs created and 0.4 million lost. It is a forecast, and Gartner counts more jobs lost than gained in 2026. Will custom software get cheaper with AI? We believe so. In our own AI builds, a system takes a fraction of the time and cost it used to, with an engineer approving every change. We have no measured multiplier to publish yet, so treat it as our experience and not as a statistic. Should companies still hire junior developers? We think so, in smaller numbers and on purpose. Experienced reviewers are the bottleneck today, and the reviewers of 2030 are juniors now. In our own work we see AI taking over junior tasks first, and we expect that squeeze to even out over the next couple of years. ## Tell us what you want to build and we will tell you what it takes. A service delivery manager replies with the roles we would assign, the monthly capacity and what the first month looks like. [Get in touch](https://azendo.co/get-in-touch/) [How it works](https://azendo.co/how-it-works/) ## Tell us about your AI project.