Estimated reading time: 4 minutes
Key takeaways:
- AI-generated code isn’t dominant, yet! Most teams use AI for ≤25% of code.
- Startups lead, enterprises lag: small teams rely heavily on AI while larger orgs are more cautious.
- Devs shift from coding to judgment. AI handles implementation as humans focus on design, review, and decisions.
The cost of writing code is quickly dropping to zero, but just how much code is AI really writing? LeadDev’s State of AI-Driven Software Releases report has some answers.
In just a few years, AI-coding assistants have gone from curiosity to commonplace. AI-generated code has become commonplace thanks to tools like Codex and Claude Code. Now the question is: will there come a day when humans no longer need to touch a keyboard to code?
Well not yet, at least according to the 400 engineers who got in touch for LeadDev’s 2026 State of AI-Driven Software Releases report.
We asked how much of your organization’s newly generated code was created at least in part with an AI-powered developer tool. Of those who could answer the question, 52% said that no more than 25% of their code was developed with the help of an AI-coding tool. Only 4% said that more than three-quarters of their created code was AI-assisted.
However, for smaller companies with fewer than 10 engineers, 45% said that more than half of their new code was generated using AI-coding tools, suggesting a real gap between more greenfield and brownfield environments.
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
The rise of AI-generated code
At cutting edge companies like Anthropic, engineers are increasingly relying on AI for coding. Boris Cherny, head of Claude Code, shared on X that he hasn’t written any code by hand for over two months. He recently shipped 49 pull requests in the space of two days – all generated by Claude.
Cherny also stated that, across the company, “pretty much 100%” of code is now AI-generated. It’s worth noting that Cherny has vested interest here. Anthropic builds AI-coding tools. If its own codebase is entirely AI-generated, it demonstrates real-world confidence in its product and creates a powerful marketing narrative.
Cherny’s experience echoes remarks made by Anthropic CEO Dario Amodei, speaking at the World Economic Forum (WEF) Annual Meeting in January. Amodei said some engineers have stopped writing code themselves and now focus on editing AI outputs instead.
The trend extends across the tech landscape. In February 2026, Spotify co-CEO Gustav Söderström revealed that the streaming service’s top developers “have not written a single line of code since December.”
As Martin Reynolds, field CTO at Harness, sees it, there’s a clear reason why organizations are going all-in on AI coding tools: to see whether teams can accomplish more with the resources they already have.
Human touch remains
While AI can clearly accelerate the coding part of the job, humans remain essential for review, debugging, and security oversight.
According to LeadDev’s report, 57% of organizations require 100% manual human review of all AI-generated code. This is because 42% of respondents distrust AI, with 38% citing security concerns and 35% worried about response accuracy and hallucinations.
“We are in a phase where humans need to communicate the restraints and then verify the output,” said Jossie Haines, a former VP of software. “Anyone who says otherwise has not dealt with the consequences of shipping subtle bugs at scale.”
Ultimately, human oversight remains essential because software does not exist in isolation. It interacts with complex systems, users, and environments, making careful review and validation indispensable.
More like this
“AI can, and increasingly will, write most of the routine code, but software engineering was never primarily about typing code,” said Haines. “Coding is just one part of the software development lifecycle, and if you speed up one part, you create pressure everywhere else.”
Agentic systems can perform substantial implementation work, such as turning a ticket into multi-file code changes, or debugging unfamiliar code, but humans still decide what to build, set constraints and architecture, and determine whether the result is correct and useful, Andrew Ambrosino, member of technical staff for OpenAI’s coding model Codex, told LeadDev.
“The responsibilities that remain uniquely human are the ones tied to context, accountability, and taste,” said Ambrosino. Deciding what should be built, understanding user needs, evaluating tradeoffs, ensuring systems behave correctly, and determining what is worth shipping are all prime examples. “These are decisions that require product judgment and real-world understanding,” Ambrosino added.
Technical decisions involve tradeoffs beyond what code gets written, shaped by constraints, organizational context, and long-term goals. AI can suggest options, but lacks full system understanding.
“The accountability needs to stay with humans,” said Diego Quiroga, senior engineering leader at Microsoft. “Even as AI becomes more capable, the role of the human shifts, rather than disappears.”
London • June 2 & 3, 2026
New activities added to LDX3 London 🎉
Devs are evolving, not going extinct
Instead of signaling a future where software no longer needs people, AI is reshaping the role of the developer, not replacing it.
“Over the next five to ten years, we expect AI systems to take on increasingly long and complex engineering tasks, but within a framework shaped by human intent and review, said Ambrosino.
“The shift we’re already seeing is from AI as a pair programmer to something that more closely resembles a teammate. Engineers increasingly define the plan, align on the approach, and then hand off implementation work to one or more agents.”
This shifts the rhythm of software development. Developers spend less time writing every line of code and more time defining problems, reviewing plans, supervising parallel AI agents, and improving product quality, reliability, and user experience.
“Engineers are still builders, but they’re more explicitly decision-makers now,” said Reynolds. “AI is handling more of the implementation, so humans are focusing on design, correctness, validation, and security.”
