AI writes most of the code now. The job that’s left is harder, not smaller.
Google says 75% of its new code is AI-generated, Microsoft 20–30%, Anthropic nearly all. That doesn't shrink engineering — it moves the job up the stack to architecture, review, and security.
The headline number keeps climbing. At Google Cloud Next this year, Sundar Pichai said 75% of new code at Google is now AI-generated and approved by an engineer — up from roughly 50% last fall and about 25% in October 2024. Microsoft puts its figure at 20–30% (Satya Nadella). Snap says 65%. Anthropic reportedly writes nearly all of its code with AI assistance. The autocomplete era is over; AI now drafts the codebase.
Here's what that statistic hides: at none of these companies is the engineering organization collapsing. The work didn't vanish. It moved up the stack.
AI is in every stage now, not just the editor
The visible role is the in-IDE co-pilot — generating boilerplate, explaining unfamiliar legacy code, and refactoring for performance. But the more consequential shift is that AI now sits across the whole lifecycle:
Testing & QA. Because code volume is exploding, automated testing is the only way to keep up — AI drafts unit and integration tests from requirements and flags which pull requests carry the highest regression risk.
DevSecOps. Vulnerability scanning at commit time, generated CI/CD and container configs, and anomaly detection on live logs to catch failures before users do.
Product & workflow. Documentation that updates itself from the actual diff, and dependency management that watches for breaking changes across services.
Gartner expects AI to touch 70% of app development work this year. The pipeline, end to end, is becoming AI-augmented.
The job that's left
Generating code was never the hard part of software. Understanding it, integrating it, and being accountable when it breaks — that was always the job. And AI hands you far more code to understand, integrate, and be accountable for.
So the role is shifting from syntax writer to architect, curator, and reviewer: someone who frames the system, judges the tradeoffs, and catches the plausible-looking mistake a model is very good at producing. That shift is hardening into new specializations:
Role
Primary focus
Core skills
AI Software Engineer
Wiring LLMs into real products via APIs
Backend, API orchestration, prompt design, system design
ML Engineer
Training and fine-tuning custom models
Python, PyTorch/TensorFlow, data pipelines
MLOps / AI DevOps
Scaling and deploying AI workloads
Kubernetes, CI/CD, IaC, cloud
AI Security Engineer
Defending against model poisoning, data leakage
Data governance, security, compliance
Our read
"X% of code is written by AI" is the most over-read number in tech. It measures the cheap part. A model going from 25% to 75% of Google's commits in eighteen months — while headcount stays roughly flat — doesn't mean engineers are three-quarters redundant. It means each engineer ships more and spends more time reviewing, integrating, and securing what the machine drafted. AI raised throughput, not unemployment. Yet.
The honest catch lives in that "yet." The comfortable story — AI writes the code, humans do the higher-bar work — assumes the higher-bar work stays human. The same labs posting these numbers are aiming AI squarely at that layer: review, planning, multi-step agents that make architectural calls. If the model climbs from drafting code to deciding the design, the "raised bar" job narrows too.
And one cost is already here. When three-quarters of your code is generated, code review and supply-chain security stop being chores and become the main event — more third-party dependencies nobody read, more confident-looking bugs, a wider attack surface. That's the exact problem security tooling is now racing to cover.
For anyone in the field, the move is the one good engineers always made: get closer to the parts AI can't be accountable for. Learn to read code faster than you write it, own the architecture, and treat "the AI wrote it" as the start of your job, not the end. The bar didn't drop. It moved — and it's higher than it looks.
Reporting from Tom's Hardware (Microsoft), Google Cloud Next 2026 (Pichai), and Gartner.
