3.1: Automation Levels L0-L3
What You'll Learn
By the end of this chapter, you will be able to:
- Define each automation level and its characteristics
- Select appropriate automation levels for different activities
- Plan transitions between automation levels
- Map automation levels to ASPICE processes
The Four-Level Model
Note: The diagram shows automation levels with Human/AI effort percentages (e.g., 75%/25% means 75% human effort, 25% AI effort).
L0: Manual
Definition
No AI assistance. Human performs all activities using traditional tools and methods.
Characteristics
| Aspect | L0 Characteristic |
|---|---|
| AI involvement | None |
| Human effort | 100% |
| Decision making | Entirely human |
| Verification | Human-performed |
| Reproducibility | Depends on individual |
When to Use L0
| Situation | Rationale |
|---|---|
| Safety-critical decisions | Human judgment required |
| Stakeholder negotiations | Interpersonal skills needed |
| Novel problem domains | No AI training data |
| Regulatory mandates | Human signature required |
| AI tool unavailable | Technical constraint |
Examples
- Safety concept development
- Stakeholder requirements elicitation
- Final release authorization
- Legal/compliance decisions
- Crisis management
L1: AI-Assisted
Definition
AI provides suggestions and recommendations. Human makes all decisions and performs primary work.
Characteristics
| Aspect | L1 Characteristic |
|---|---|
| AI involvement | 25% (suggestions) |
| Human effort | 75% (decisions + work) |
| Decision making | Human with AI input |
| Verification | Human judges AI suggestions |
| Reproducibility | More consistent with AI support |
AI Role at L1
The following diagram illustrates the L1 workflow where the human drives the process and AI provides suggestions that the engineer accepts or rejects at each step.
When to Use L1
| Situation | Rationale |
|---|---|
| Complex domain knowledge | AI lacks full context |
| Judgment-intensive work | Human expertise critical |
| High-stakes decisions | Need human accountability |
| Early AI adoption | Building trust incrementally |
| Ambiguous requirements | Need human interpretation |
Examples
- Requirements quality analysis (AI flags issues, human resolves)
- Architecture review (AI suggests patterns, human decides)
- Code completion (AI suggests, human accepts/rejects)
- Security threat analysis (AI identifies, human prioritizes)
L2: High Automation
Definition
AI generates significant output. Human reviews, validates, and approves.
Characteristics
| Aspect | L2 Characteristic |
|---|---|
| AI involvement | 70% (generation) |
| Human effort | 30% (review/approval) |
| Decision making | AI proposes, human approves |
| Verification | Human reviews AI output |
| Reproducibility | High (AI consistent) |
AI Role at L2
At L2, the AI generates complete artifacts and the human's role shifts to reviewing and approving the output before it proceeds.
When to Use L2
| Situation | Rationale |
|---|---|
| Routine generation tasks | High volume, consistent format |
| Well-defined outputs | AI can be evaluated objectively |
| Proven AI capability | AI has demonstrated accuracy |
| Efficient human review | Outputs can be quickly assessed |
| Time-sensitive work | AI speed beneficial |
Examples
- Unit test generation (AI generates, human reviews)
- Code review (AI first-pass, human approval)
- Documentation generation (AI drafts, human edits)
- Traceability linking (AI suggests, human confirms)
- Bug triage (AI categorizes, human verifies)
L3: Full Automation
Definition
AI executes autonomously. Human monitors results and intervenes on exceptions.
Characteristics
| Aspect | L3 Characteristic |
|---|---|
| AI involvement | 90% (execution) |
| Human effort | 10% (monitoring) |
| Decision making | AI decides routine cases |
| Verification | Automated with human spot-checks |
| Reproducibility | Very high |
AI Role at L3
At L3, the AI executes autonomously while the human monitors aggregate results and intervenes only when exceptions or anomalies arise.
When to Use L3
| Situation | Rationale |
|---|---|
| Highly routine operations | Predictable, repetitive |
| Automated quality gates | Objective pass/fail criteria |
| Low-risk activities | Errors are recoverable |
| High volume | Manual infeasible |
| Proven reliability | >99% accuracy demonstrated |
Examples
- CI/CD pipeline execution
- Static analysis running
- Format checking
- Build automation
- Regression test execution
- Deployment to staging environments
Level Selection Guide
Decision Matrix
| Factor | L0-L1 | L2 | L3 |
|---|---|---|---|
| Risk if AI errors | High | Medium | Low |
| Human judgment needed | High | Medium | Low |
| Task repeatability | Low | Medium | High |
| AI proven accuracy | Low | High | Very High |
| Volume of work | Any | Medium-High | Very High |
| Verification difficulty | Easy | Moderate | Automated |
ASPICE Process Mapping
| Process | Recommended Level | Rationale |
|---|---|---|
| SYS.1 | L0-L1 | Stakeholder interaction |
| SYS.2 | L1 | Consistency checking |
| SYS.3 | L1 | Pattern suggestions |
| SWE.1 | L1-L2 | Derived requirements |
| SWE.2 | L1 | Architecture patterns |
| SWE.3 | L2 | Code generation |
| SWE.4 | L2-L3 | Test generation/execution |
| SWE.5 | L2 | Integration testing |
| SWE.6 | L1-L2 | Qualification testing |
| SUP.8 | L3 | Configuration management |
| MAN.3 | L1 | Project analytics |
Level Transitions
L0 → L1 Transition
Prerequisites:
- AI tool evaluated
- Quality baseline established
- Team trained
Process:
- Run AI in shadow mode
- Compare AI suggestions to human work
- Measure agreement rate
- Enable AI suggestions when agreement >70%
L1 → L2 Transition
Prerequisites:
-
90% acceptance of AI suggestions
- Review process defined
- Rollback capability in place
Process:
- AI generates for subset of work
- Track acceptance rate
- Retrain on rejected outputs
- Full L2 when acceptance >80%
L2 → L3 Transition
Prerequisites:
-
99% accuracy demonstrated
- Automated validation in place
- Monitoring operational
Process:
- Reduce review frequency
- Monitor for anomalies
- Maintain audit trail
- Human review on metrics triggers
Summary
The four automation levels provide a structured approach to AI integration:
| Level | Human | AI | Decision Authority |
|---|---|---|---|
| L0 | 100% | 0% | Human only |
| L1 | 75% | 25% | Human with AI input |
| L2 | 30% | 70% | AI proposes, human approves |
| L3 | 10% | 90% | AI executes, human monitors |
Selection based on: risk, judgment needs, repeatability, proven accuracy, and volume.