How much code AI already writes
The loudest figure came from Google. On 22 April 2026 its chief executive, Sundar Pichai, reported that AI creates 75% of all new code at Google, and engineers approve it. In autumn 2025 the share was 50%, and in October 2024 more than a quarter. A complex code migration that agents and engineers did together was finished six times faster than engineers alone managed a year earlier.
Almost everyone uses the tools. In the Stack Overflow 2025 survey, 51% of professional developers work with AI every day. In the DORA 2025 study, which surveyed almost 5,000 professionals, 90% use AI, and more than 80% say it raised their productivity.
An important caveat: 75% of code is not 75% of the work. Writing code is one part of a developer’s task. The rest is to understand what the business needs, design the system, check changes, release them and answer for failures. Reread the second half of Google’s sentence: engineers approve the code. A person still decides whether a change reaches the live system.
Where AI speeds developers up and where it slows them down
Surveys show feelings, not time. The nonprofit research organization METR measured the time. In a randomized trial in 2025, 16 experienced open-source developers solved 246 real tasks, and for each task it was decided at random whether AI was allowed. With AI the tasks took 19% longer. Before the start the participants expected a 24% speedup, and after the trial they still believed AI had sped them up by 20%.
In February 2026 METR repeated the measurement with tools from late 2025: 57 developers, more than 800 tasks. The estimate turned into a speedup: 18% for participants of the first trial and 4% for new ones. But METR itself calls this signal unreliable. Many developers refused to take part because they did not want to work without AI, and 30 to 50% of those asked admitted that they kept out of the trial tasks they did not want to do by hand. The researchers’ conclusion: developers with AI are now most likely faster than in early 2025, but by how much is not yet known.
Two things follow for a company. First, the team’s self-assessment overstates the effect: people are sure of a speedup even where there is none. Second, the effect depends on the task and on how the work is organized. The DORA report puts it this way: AI amplifies an organization’s strengths and weaknesses. If the team has tests and code review, AI speeds it up. If not, it produces errors faster.
When will AI replace programmers: the employment forecast

The US Bureau of Labor Statistics publishes a regular forecast for the profession. By its update of 27 August 2026, there are about 1.9 million software developers together with testers in the US, and by 2035 their number should grow by 10%, faster than the average for all occupations. The Bureau expects about 106 thousand openings a year. One reason for the growth: companies need more software for AI, robots and automation.
For programmers in the narrow sense, those who write and test code from a ready specification, the forecast is the opposite: minus 7% by 2035, from 110,800 jobs today. The Bureau names the reason directly: companies will automate repetitive programming tasks, including with AI, and part of the complex work will move to developers.
This is the honest answer to the question in the title. AI takes the work of “writing code from a description” and does not take the work of “deciding what to build and answering for the result”. The profession does not disappear but shifts: fewer people write code by hand, more set tasks, check and answer for the system. A similar analysis for another profession: will AI replace a lawyer.
What is happening to the hiring of junior developers
Young professionals felt the changes first. Stanford researchers Erik Brynjolfsson, Bharat Chandar and Ruyu Chen studied the payroll records of millions of US workers from the ADP system. In the August 2026 version, employment of workers aged 22 to 25 in the occupations most exposed to AI, including software development, is 19% lower than it would be with the same growth as their peers in less exposed occupations. Experienced workers show no such gap.
For an executive the mechanism matters: the gap came not from layoffs but from hiring fewer young people. The authors are careful and call their findings early descriptive signals, not a proven cause. But the direction is clear: companies give AI the work on which beginners used to learn.
In Russia the market has changed too, though for other reasons. According to an hh.ru report published on 16 March 2026, the hh-index in IT for 2025 was 14.2, meaning it doubled in a year, after holding between 7.3 and 7.7 for three years. hh.ru is Russia’s largest job site, and its index shows the number of job seekers per opening. The number of active job seekers grew by 12.6%, and hh.ru explains this by new specialists entering the market, including after courses. Hiring a developer is now easier than three years ago.
This creates a risk that is easy to miss. If a company stops hiring beginners, in five years it will have nobody to grow into experienced staff. An experienced developer grows from a beginner who worked for several years next to strong colleagues. It is wiser to hire fewer beginners but teach them from day one to set tasks for AI and check its code.
What this means for the IT budget
Cutting the team in proportion to AI’s share of code is a mistake. If AI writes three quarters of the code, you do not need three quarters fewer developers: review, design and responsibility for failures remain, and there is more code than the team can read.
Money moves between budget lines. Routine work gets cheaper: function drafts, tests, migration to new library versions, documentation. New costs appear: subscriptions to AI tools for each developer, experienced people’s time to review others’ code, and security checks. If the source code must not go to an outside model vendor, a model on your own servers is added to them.
Measure the effect by tasks, not by lines of code or by the team’s feelings: how many tasks were closed per month before AI tools, how many after, and how many errors were found after release. The METR measurement shows why feelings cannot be trusted. We covered how to review code written by an agent on the page about agents for programming, and how to set a goal in numbers in the article on what to expect from deploying AI.
Who reviews AI code and who answers for it

Responsibility for AI code stays with the company. A model does not sign a contract, does not answer the phone at night during an outage and does not answer to a client. So every change must have a person who accepted it and answers for it. At Google this is what engineer approval means.
Developers themselves trust AI with caution. By DORA 2025, 30% of professionals trust AI results a little or not at all. In the Stack Overflow survey, 75% of developers said that in the future they will still go to a person for help when they do not trust an AI answerAI answers in search. AI answers in search are short summaries that a search engine writes with a language model above the list of links, built from several sites.. This is a healthy position: AI code is reviewed as strictly as the code of a new employee.
Before the team starts working with agents, settle three questions. Who reviews each change, and in what time. Where the source code and customer data go: data about Russian citizens is stored in databases in Russia under 152-FZ, Russia’s personal data law, and the source code is not sent to an outside model vendor without the company’s decision. Who owns the code a contractor writes with AI: the contract must say this as clearly as for code written by people.
What an IT department can give to AI agents now
The guideline is simple: give the agent work whose result can be checked by a test or a comparison, and where an error is cheap.
1. Drafts of tests and standard functions from the task description. A person approves the test, the agent writes the code.
2. Migration of code to new versions of libraries and systems. This is exactly the work where Google got its sixfold speedup.
3. Documentation for existing code and descriptions of changes for colleagues.
4. The first analysis of an error from the log: the agent finds the likely cause and suggests a fix, and the developer decides whether to accept it.
5. The first review of changes before a human review: obvious errors, missing tests, breaks of formatting rules.
6. Answers to standard staff questions to IT support from the knowledge baseAI knowledge base. An AI knowledge base is a set of selected and prepared company documents that an AI assistant searches for answers: rules, instructions, a product catalog and answers to common questions..
7. Internal tools and prototypes: a request form, a reference list, a simple report. What services can build without a programmer is covered in our review of AI for building apps.
What stays with people
Deciding what to build and why. An agent does not know which process in sales or in the warehouse is worth automating first and how much money it will bring.
Architecture: how the company’s systems connect, where customer data is stored and what happens when the load grows. Accepting changes and answering for failures in the live system. Access and security: what rights to give an AI agent and which keys not to hand over.
Knowledge of the business. The IT department makes sure systems work, while decisions on how sales or support work will change need knowledge of those processes. Why this is hard for an internal team is covered in the article about deploying AI on your own.
The conclusion for hiring: hire fewer people who only write code and more who can set a task, check the result and answer for the system. Which professions AI will not replace, apart from development, we have collected separately.



