I’ve been using AI code assistants daily since early 2024. GitHub Copilot, Cursor, Claude, ChatGPT — I’ve run them all through real production workflows, not toy examples. After two years, the hype has settled enough to give an honest assessment of where these tools genuinely help and where they waste your time.

What Actually Works

Boilerplate generation. This is where AI assistants earn their keep. Writing Express route handlers, React component skeletons, database migration files, test setups — repetitive structural code that follows predictable patterns. I estimate AI handles about 70% of this work correctly on the first attempt, and the remaining 30% needs minor edits rather than rewrites.

Explaining unfamiliar code. When I’m diving into a new codebase, asking an AI to explain a complex function is faster than reading it line by line. The explanations aren’t always perfect, but they give me enough context to understand the intent, which I can then verify against the implementation.

Regex and shell commands. I will never memorise regex syntax, and AI assistants have made that acceptable. Describing what I want to match in plain English and getting a working regex back is genuinely transformative for my workflow. Same goes for obscure find, awk, and sed commands.

Test generation. Given a function with clear inputs and outputs, AI assistants generate reasonable unit tests. They’re good at covering happy paths and common edge cases. They miss domain-specific edge cases, but the generated tests serve as a solid starting point.

What Doesn’t Work

Complex architecture decisions. AI assistants will confidently suggest architectural patterns that are inappropriate for your specific context. They don’t understand your team’s skill level, your deployment constraints, or your maintenance budget. I’ve seen junior developers follow AI-suggested architectures that introduced unnecessary complexity because the AI defaulted to “enterprise patterns” for a simple CRUD app.

Debugging production issues. When something breaks in production, the context required to diagnose the problem — deployment configuration, infrastructure state, recent changes, traffic patterns — is almost never available to the AI. It can suggest generic debugging steps, but it can’t reason about your specific system’s behaviour.

Security-sensitive code. AI assistants occasionally generate code with security vulnerabilities. Improper input validation, SQL injection vectors, missing authentication checks — I’ve caught all of these in AI-generated code. Never ship security-critical code without human review, regardless of how confident the AI seems.

Businesses that are serious about integrating AI into their development process often work with an AI consultancy to establish proper review workflows and guardrails. It’s not enough to just hand developers a tool — you need processes around how generated code gets validated.

The Productivity Question

The common claim is that AI assistants make developers “10x more productive.” That hasn’t been my experience. My rough estimate is a 15-30% productivity increase for experienced developers, concentrated in specific tasks:

• Writing new code from scratch: 20-30% faster
• Refactoring existing code: 10-15% faster
• Writing tests: 30-40% faster
• Debugging: minimal improvement
• Code review: slightly slower (you’re now reviewing AI output too)

For junior developers, the picture is more complicated. AI assistants can accelerate output but may slow learning. If you accept every suggestion without understanding it, you’re building on a foundation you can’t maintain.

Tool-Specific Notes

GitHub Copilot remains the best option for inline completions while typing. It’s deeply integrated into VS Code and understands file context well. The chat feature is decent but not best-in-class.

Cursor offers the best “AI-native” editor experience. The ability to reference specific files and apply multi-file edits is ahead of the competition. It’s my primary editor for greenfield projects.

Claude (via API or direct) produces the highest quality code for complex, multi-step tasks. When I need to design a module from scratch with proper error handling and types, Claude consistently outperforms the alternatives. The trade-off is speed — it’s slower than inline completion tools.

Practical Recommendations

Use AI assistants as a first draft generator, not a final answer provider. Write a comment describing what you want, let the AI generate a solution, then review and refine it. This workflow captures most of the productivity gains while maintaining code quality.

Invest time in learning how to write effective prompts. The difference between a vague request and a specific one with context is enormous. Include type signatures, describe edge cases, and specify error handling expectations.

Don’t fight the tool on tasks it’s bad at. If the AI generates something wrong twice, write it yourself. The time spent iterating on bad suggestions exceeds the time to write the code manually.

AI code assistants are a permanent part of professional development now. They’re useful, imperfect, and improving. Treat them as capable but unreliable colleagues who need supervision, and you’ll get the most value from them.