Configuring an Autosar NvM (Non-Volatile Memory) stack involves coordinating 9 BSW modules with precise cross-references. We evaluated whether AI could handle this complexity while maintaining the referential integrity that typically requires manual verification.
Results: 7 of 9 modules fully configured in ~1 hour (versus 4-8 hours when done manually), with the AI appropriately stopping at hardware and vendor integration boundaries.
A Software Component requiring persistent storage triggers configuration across the entire memory stack:
Memory Stack Layers:
Integration Layers:
Typical engineer time: 4-8 hours
Common errors: Mismatched block IDs, incorrect device references, size misalignments
We used a reasoning model (Gemini 3 Pro) to decompose the requirement into 10 sequential prompts, one per BSW module plus a final validation prompt. The model identified the correct dependency order to prevent forward references.
After each configuration step, Wings validated against the Autosar schema. The AI used this feedback to iterate and correct issues before proceeding.
| Module | Status | Key Configuration | Cost |
|---|---|---|---|
| Fee | Complete | Block 2, 64 bytes | $0.62 |
| NvM Block | Complete | CRC16, retry logic, references | $4.09 |
| NvM Common | Complete | API class, queues, 10ms period | $1.93 |
| MemIf | Complete | Device mapping verified | $0.13 |
| RTE | Partial | Module instance created, implementation ref missing | $1.65 |
| Dem | Complete | 3 diagnostic events linked | $8.85 |
| BswM | Complete | Startup/shutdown coordination | $13.99 |
| EcuM | Complete | Initialization sequence configured | $2.23 |
| Os | Complete | Task and 10ms alarm created | $2.54 |
| Fls | Partial | Requires hardware-specific parameters | $0.30 |
| Total | 7/9 | ~1 hour execution time | $37.03 |
The AI maintained consistency across modules configured in isolation:
Fee-NvM Matching:
Device References:
This type of consistency is difficult to maintain manually across separate .arxml files and is typically caught only during integration testing.
The AI used validation feedback to iterate on the NvM block descriptor:
Cost of getting it right: $4.09 (121k tokens with context compression)
Manual approach: Engineers typically configure everything first, then spend 30-60 minutes debugging validation errors at the end.
RTE Module: AI stopped when it couldn't find BSW implementation definitions, documenting exactly what the vendor integration manual should provide.
Fls Module: AI refused to configure hardware-specific parameters (flash sectors, addresses, page sizes) without microcontroller specifications. It generated a detailed requirements report instead.
These stopping points demonstrate domain understanding rather than hallucination.
Post-execution validation identified:
Strengths:
Gaps:
Assessment: 70-80% complete in 1 hour. Remaining work requires hardware specifications and vendor integration details that were unavailable to the AI.
Traditional approach:
AI-assisted approach:
Savings: $213-938 per configuration (54-78% reduction)
For a project with 10 SWCs requiring NvM access:
| Approach | Total Cost |
|---|---|
| Traditional | $4,000-12,000 |
| AI-Assisted | $1,870-2,620 |
| Project Savings | $2,130-9,380 |
The division of labor puts AI on tedious configuration tasks and keeps engineers focused on decisions that require domain expertise and judgment.
This evaluation demonstrates that AI can handle complex, multi-module Autosar configuration when provided with:
The AI maintained referential integrity across isolated configurations and recognized its boundaries appropriately. The resulting 75% time reduction comes from automating the mechanical aspects of BSW configuration while preserving engineer control over hardware and architectural decisions.
Limitations: This represents a single configuration scenario. Different BSW stacks, Autosar versions, and project constraints may yield different results.
Interested in seeing how Wings handles your Autosar project?
Want access to the complete prompt sequences and AI responses from this case study? Contact us to discuss your specific use case.
Based on Wings v0.21.0 usage in February 2025. All metrics and validation results are from actual project execution.
