Software Engineer.
AI automates boilerplate and junior-level coding tasks; engineers who architect systems and direct AI tools effectively are increasingly valuable.
High transformation, medium risk.
The pattern emerging in software engineering mirrors what happened in other knowledge work professions. AI is very good at the structured, repeatable parts of the job, and not yet good at the judgment-intensive parts. The job at the junior and mid level has always been more structured than the profession liked to admit.
4 shifts already visible in the data, in order of magnitude.
AI code completion is now standard infrastructure.
Tools like GitHub Copilot, Cursor, and Claude generate contextually relevant code completions, full functions, and entire modules from natural language. Developers using Copilot complete discrete tasks 55% faster on average.
The entry-level market has materially contracted.
Tasks that defined junior engineer work — writing CRUD endpoints, scaffolding test files, building simple UIs — are precisely the tasks AI handles well. Entry-level technical hiring at major tech companies declined significantly between 2023 and 2025.
Non-engineers are building software through prompting.
Internal tools, dashboards, and simple automations that previously required engineering resources are now being built by product managers, analysts, and operators using Cursor and Claude. This compresses demand for routine application development.
Autonomous AI software agents are arriving faster than expected.
Systems from Cognition, Google, and others can take a GitHub issue, write code to address it, run tests, and open a pull request. Current success rates on realistic engineering benchmarks are still well below human, but the trajectory makes meaningful contribution to maintenance plausible within a few years.
What the leaders are doing.
| № | Company | Sector | What they are doing | Year | Source |
|---|---|---|---|---|---|
| 01 | Microsoft / GitHub | Developer Tools | GitHub Copilot surpassed 4.7 million paid subscribers by early 2026, up 75% year-over-year. Microsoft reports 55% faster task completion in controlled trials, and has embedded Copilot across VS Code, GitHub Actions, and Azure DevOps. | 2026 | github.blog ↗ |
| 02 | Technology | CEO Sundar Pichai reported in 2024 that AI tools generate more than 25% of new Google code, reviewed and accepted by human engineers — a figure that has continued rising with broader Gemini Code Assist adoption internally. | 2025 | theverge.com ↗ | |
| 03 | Cognition AI | AI / Developer Tools | Devin is an autonomous AI software agent that handles end-to-end engineering tasks: scaffolding projects, debugging, writing tests, and opening pull requests. Enterprise pilots are active for well-specified maintenance and feature work. | 2025 | cognition.ai ↗ |
What is declining, growing, emerging.
- 01Writing boilerplate code, scaffolding, and repetitive implementation patterns from scratch
- 02Manual unit test and integration test authoring for standard, well-understood code paths
- 03Basic debugging using stack trace analysis for well-understood error types
- 04Writing developer documentation and code comments from scratch
- 01System architecture and high-level design — trade-off analysis that requires human judgment
- 02AI output review and validation: reading generated code for correctness, security flaws, and edge cases
- 03Cross-functional communication between engineering, product, and business stakeholders
- 04Debugging complex distributed systems, race conditions, and performance issues beyond AI reasoning ability
- 01Prompt engineering for code generation — writing specifications that produce reliable AI-generated output
- 02LLM application development: building systems where language models are core architectural components
- 03AI agent evaluation and orchestration — designing and assessing multi-step autonomous engineering workflows