AI värde utvecklingsfaser visar asymmetrisk distribution i Humboldt University survey (65 developers + literature review): 79% använder GenAI dagligen, 85% rapporterar högst värde i Design/Implementation, 75% i Operation/Maintenance, men bara 18% i Planning, 43% i Requirements Analysis. För dig som projektledare betyder detta: GenAI transformerar inte hela SDLC uniformt. Value concentrates där tasks are syntactically structured (boilerplate, documentation), disappears där judgment/context critical (planning, requirements). 72% halve time för boilerplate, 69% för documentation, men debugging visar weakest gains + 6% report increased effort. Value shift från routine coding → specification quality, architectural reasoning, oversight. Governance maturing (66% har formal/informal guidelines) men risks (skill erosion, technical debt, uncritical adoption) require robust human-in-the-loop mechanisms.
Survey findings: 79% daily use, browser-based LLMs dominate
Demographics (n=65):
- 30 professional developers, 4 product owners, 1 team lead
- Young age structure men 10 participants >10 years experience
- 42% Scrum/Kanban, 34% no defined process, 17% hybrid, 8% traditional
Daily usage:
- 79% use AI tools ≥1x/day (routine component av professional practice)
Tool preferences:
- ChatGPT: 77% (most adopted)
- Google Gemini/Bard: 52%
- GitHub Copilot: 48% (IDE-integrated + chat)
- JetBrains AI Assistant: 11%
- Tabnine: 3%
- Amazon CodeWhisperer: 2%
Pattern: Developers favor dialog-based general-purpose LLMs över specialized IDE-integrated alternatives
För projektledare: Browser-based interaction preferred över IDE plugins. Teams likely using ChatGPT for broader problem-solving beyond pure code generation.
SDLC phase impact: Massive disparity
Perceived greatest utility (multiple choice):
HIGH:
- Design & Implementation: 85%
- Operation & Maintenance: 75%
LOW:
- Requirements Analysis: 43%
- Testing & Integration: ~37%
- Planning: 18% ⚠️
Critical insight: None av Product Owners + only 5% av >10yr experience developers report impact i Planning phase
För projektledare: AI strongest där work är syntactically structured. Planning requires strategic thinking, stakeholder alignment, risk assessment – areas där AI currently weak. Don’t expect AI substitute project managers’ judgment.
Task-level time savings: Boilerplate + docs highest
Estimated time required med GenAI (vs utan AI):
Strongest effects (≥50% reduction):
- Boilerplate code: 72% halve time
- Documentation: 69% halve time
- New code: 37% report ~75% time
- Unit tests: Substantial gains (men 28% “cannot assess”)
Moderate/Weak:
- Understanding unfamiliar/legacy code: Polarized (29% extreme speedup <25%, men 11% no change)
- Debugging: Weakest gains + 6% report INCREASED effort ⚠️
För projektledare: Automation benefits concentrated i repetitive, standardized tasks. Complex cognitive work (debugging) gains little eller worsens. Measure productivity by task category, not blanket “AI boost”.
Value shift: Routine → Judgment architecture
Literature findings + qualitative responses converge:
FROM: Routine coding (boilerplate, documentation, repetitive tasks) TO:
- Specification quality
- Architectural reasoning
- Critical validation
- Governance oversight
Qualitative themes:
Efficiency dominant: “Time savings through automation repetitive tasks” “Sparring partner accelerates learning” “Reduces toil, allows focus on architecture/complex problem solving”
Concerns substantial: “Uncritical adoption AI-generated code” “Loss deep technical understanding” “Blind copying – outputs appear correct but contain logical/security flaws” “Cognitive offloading: if AI shortcut not support tool, learning curves deteriorate”
Role transformation expected: “Shift från pure code writing → architectural thinking, requirement formulation, QA, orchestration AI systems” “Polarization: junior roles decline, senior/architect competencies more valuable” “New hybrid roles: AI supervisors”
För projektledare: GenAI inte eliminating expertise need – shifting WHERE expertise applied. Senior talent becomes MORE valuable (validation, architecture, strategic decisions). Junior pipeline MAY shrink.
Governance maturity: 66% har guidelines men gaps remain
Organizational AI policies:
- 40%: Official tools + formal AI policy
- 26%: Informal recommendations
- Total structured guidance: 66%
Men:
- 20%: Unaware av existing policies
- 8%: No regulation exists
- Small proportions: Forbidden, no response, other
För projektledare: Majority organizations addressing governance, but non-negligible segment exhibits regulatory ambiguity eller limited transparency. Implement:
- Clear usage policies
- Approved tool lists
- Security/compliance guidelines
- Review obligations för AI outputs
Risks: Skill erosion, technical debt, security
Short-term:
- AI-generated code frequently erroneous (avg ~10 min correction effort per defect)
- Significant proportion outputs requires post-editing
- Superficial quality control: Developers accept snippets with limited scrutiny
- Perceived ≠ objective productivity: Risk overestimating efficiency gains
- Security vulnerabilities: Prompt injection attacks, data contamination, unintended code execution
Long-term:
- Skill erosion: Continuous reliance erodes deep technical knowledge
- Technical debt: Auto-generated code introduces opaque dependencies, architectural inconsistencies
- Syntactically correct ≠ semantically sound: Subtle security weaknesses
- Vendor lock-in risk
- Reduced human interaction: Knowledge silos, diminished interpersonal trust
Organizational level: Quality assurance, accountability, compliance, traceability become MORE salient när AI artifacts incorporated into production
För projektledare: Formalize review obligations, clarify responsibility för AI-assisted outputs, establish human-in-the-loop validation mandatory.
Early SDLC phases: Why planning shows lowest benefit
Planning phase (18% perceived impact):
- Supports artifacts: scope, timelines, milestones, role descriptions, risk assessments
- Market analyses, competitive information, product goals
- Cost estimation, story point approximation
Men empirical evidence weak för strategic planning.
Requirements Analysis (43%):
- Drafts user stories, use cases
- Preliminary architectural drafts
- Identifies inconsistencies, ambiguities
Better than Planning men still below Implementation.
Explanation: Early phases require:
- Ambiguity tolerance
- Stakeholder negotiation
- Strategic judgment
- Context-heavy understanding
GenAI excels syntactic tasks, struggles semantic/strategic reasoning.
För projektledare: AI useful för documentation, template generation i planning. NOT substitute för strategic thinking, stakeholder alignment, risk assessment judgment calls.
Fem praktiska implementation insights
1. Phase-appropriate AI deployment: Don’t force AI across all SDLC phases uniformly High value: Implementation, documentation, maintenance Low value: Planning, requirements (strategic judgment remains human domain)
2. Task-level productivity tracking: Measure by task category (boilerplate, docs, debugging, new features) NOT blanket “AI boost percentage” Debugging may INCREASE effort despite promises
3. Human-in-the-loop mandatory: Every AI output requires validation Formalize review obligations Coding accounts only 10-15% total development time (BCG report) – automation limited impact utan broader workflow optimization
4. Skill preservation programs: Junior developers extra vulnerable skill erosion Structured training ensuring foundational knowledge Balance AI efficiency gains mot long-term capability preservation
5. Governance frameworks: Clear policies on tool usage, data security, output accountability 66% har guidelines – be in that majority Address compliance, traceability, review standards
Bottom line
AI värde utvecklingsfaser asymmetriskt distributed: 79% developers use GenAI daily, 85% högst värde Design/Implementation, 75% Operation/Maintenance, men 18% Planning, 43% Requirements. Browser-based LLMs (ChatGPT 77%) preferred över IDE plugins (Copilot 48%). Task-level: 72% halve time boilerplate, 69% documentation, men debugging weakest (6% increased effort). Value shift från routine coding → specification quality, architectural reasoning, oversight. Governance maturing (66% formal/informal guidelines) men gaps remain (20% unaware, 8% no regulation). Risks: skill erosion, technical debt, uncritical adoption, security vulnerabilities. Early SDLC phases show lowest benefits (strategic judgment AI-resistant). Literature + qualitative: role transformation från code writing → architectural thinking, validation, AI orchestration. Senior expertise becomes MORE valuable. Implementation requires phase-appropriate deployment, task-level tracking, mandatory human-in-the-loop, skill preservation, robust governance.
Källa: “The State of Generative AI in Software Development: Insights from Literature and a Developer Survey” av Vincent Gurgul, Robin Gubela & Stefan Lessmann, Humboldt-Universität zu Berlin, publicerad 19 mars 2026.