Software Engineering with LLMs in 2025: Reality Check

Gergely Orosz presents a reality check on AI in software engineering at LeadDev, cutting through the hype to examine what’s actually working.

The Current State

What’s Real

• AI coding assistants are genuinely useful
• Productivity gains exist but vary widely
• Adoption is accelerating across the industry

What’s Hype

• “AI will replace developers” narratives
• 10x productivity claims
• Autonomous coding agents (mostly)

Productivity: The Numbers

Gergely examines actual productivity data:

Studies Show

• 20-40% improvement on specific tasks
• Highly variable by task type
• Senior developers often benefit more
• Learning curve is real

What’s Not Measured

• Code quality over time
• Maintenance burden
• Technical debt accumulation
• Team dynamics changes

How Teams Are Actually Using AI

Common Patterns

1. Code completion: Most widespread use
2. Documentation: Writing and understanding
3. Testing: Generating test cases
4. Debugging: Explaining errors
5. Learning: Understanding new codebases

Less Common (But Growing)

• Code review assistance
• Architecture suggestions
• Refactoring recommendations

The Adoption Curve

Early Adopters

• Individual developers experimenting
• Startups moving fast
• Tech-forward enterprises

Mainstream

• Enterprise rollouts beginning
• Policy frameworks emerging
• Training programs developing

Laggards

• Regulated industries (banking, healthcare)
• Security-conscious organizations
• Legacy-heavy environments

Challenges in Practice

Technical

• Context limitations
• Hallucinations in code
• Integration with existing tools
• Security concerns

Organizational

• Policy uncertainty
• Training needs
• Measuring ROI
• Change management

What’s Working

Individual Level

• Faster boilerplate writing
• Better documentation
• Quicker learning
• Reduced context switching

Team Level

• Shared prompts and patterns
• Consistent code style
• Faster onboarding
• Knowledge capture

What’s Not Working

Overpromised

• Fully autonomous coding
• Replacing junior developers
• Eliminating code review
• Perfect code generation

Underdelivered

• Complex refactoring
• Architectural decisions
• Cross-system understanding
• Long-term maintenance

Recommendations

For Individual Developers

1. Experiment actively but critically
2. Learn prompt engineering basics
3. Verify AI output always
4. Share learnings with team

For Engineering Leaders

1. Set realistic expectations
2. Invest in training
3. Measure carefully
4. Update processes gradually

For Organizations

1. Develop clear policies
2. Address security concerns
3. Plan for change management
4. Budget appropriately

Looking Ahead

2025-2026 Predictions

• Better tool integration
• More specialized models
• Clearer ROI metrics
• Mature best practices

Uncertainties

• Regulatory landscape
• Model capabilities ceiling
• Economic factors
• Competitive dynamics

Key Takeaway

“AI tools are genuinely useful, but they’re tools. The fundamentals of good software engineering haven’t changed.”