Review-ready drafts
Requirements, design, test specifications, traceability, review, and quality records are review-ready, high-quality drafts, not completed or approved deliverables.
Consulting
Automotive Software & A-SPICE
AI Coding Agent Consulting
We apply AI coding agents to SDV and automotive product development under A-SPICE, ISO 26262, SOTIF, and customer processes.AI does not replace professional judgment or guarantee A-SPICE compliance, assessment, certification, ISO 26262 or SOTIF safety decisions, or safety-case approval.Documents remain review-ready, high-quality drafts until people verify, revise, approve, and baseline them.Codex and Claude Code are examples in a vendor- and IDE-neutral model, and the workflow remains fail closed when evidence or authority is missing.
Boundaries
Define the boundary between AI assistance and A-SPICE accountability
Support faster SDV, OTA, ADAS, and connected-service development while accountable people retain compliance, safety, evidence acceptance, baseline, release, and assessor decisions.
Human professional judgment
Process owners, engineers, quality and safety staff, assessors, and final approvers retain process tailoring, architecture, ISO 26262 and SOTIF safety judgments, approvals, baselines, evidence acceptance, issue closure, and release accountability.
Evidence from real execution
Never invent unexecuted test results, reviews, or approvals. Preserve provenance and stop under a fail-closed rule when evidence is missing.
Three Support Areas
Connect document drafts, implementation, and real verification
The engagement distinguishes draft, implementation, and accepted evidence while connecting requirements through code, tests, and traceability.
Document deliverable drafts
Create review-ready drafts for requirements, architecture, detailed design, interfaces, test specifications, traceability, reviews, and quality records. Human review, revision, approval, and configuration baselining are required.
Source and test code development
Develop production-quality source, unit tests, and integration tests from approved requirements, design intent, interfaces, and coding rules. Human review, real builds, tests, security, and project gates remain mandatory.
Actual tool execution and evidence packaging
Run static analysis, unit, integration, and regression tests with real tools; collect results, inspect traceability gaps, and preserve provenance, execution time, versions, commit SHA, and tool versions.
A-SPICE Process Map
Map bounded assistance across SWE.1 through SWE.6
Project tailoring determines inputs and completion criteria.AI assists drafts, implementation, execution, and gap detection; people accept results and make final decisions.
SWE.1 Software Requirements Analysis
Draft refined requirements, acceptance criteria, consistency questions, and trace links for accountable review and confirmation.
SWE.2 Software Architectural Design
Draft components, interfaces, alternatives, impacts, and review questions; architects confirm technical and safety decisions.
SWE.3 Software Detailed Design and Unit Construction
Develop source and unit tests aligned to detailed design intent, coding rules, requirements, and interfaces, then verify through review and real builds.
SWE.4 Software Unit Verification
Execute static analysis and unit tests, collect results and versions, and present them for human acceptance against approved criteria.
SWE.5 Software Integration and Integration Test
Draft integration sequencing and interface tests, then automate real execution, result collection, and defect linkage.
SWE.6 Software Qualification Test
Execute requirement-based and regression tests and inspect result and requirement-to-test traceability gaps for approval review.
AI Support vs. Human Accountability
Separate assistance from accountable decisions
SUP.1, SUP.8, SUP.9, and SUP.10 can use draft checklists, records, impact candidates, and evidence bundles, while final assurance, approval, baselines, issue closure, and change decisions remain human responsibilities.
AI: drafts and traceability
Draft documents, checklists, and trace links and identify missing or inconsistent candidates for review.
AI: code, tests, and execution
Develop code and tests, run permitted tools, organize results, and report traceability gaps without inventing evidence.
People: engineering and safety decisions
Confirm requirements and architecture, assess safety and risk, review code, and accept verification results.
People: baselines and assessment
Own tailoring, approvals, configuration baselines, release, evidence acceptance, issue closure, and assessor responses.
Agent, Skill & Human in the Loop
Design repeatable procedures, separated roles, and approval gates
Connect vendor- and IDE-neutral Agents, Subagents, Skills, Human in the Loop, and Guardrails to existing ALM, configuration, CI/CD, verification tools, internal AI platforms, and Local LLM environments.
Skill
Packages repeatable procedures and tool-use rules for each A-SPICE activity, including inputs, checks, commands, evidence, and stop conditions.
Subagent
Separates requirements, design, implementation, testing, traceability, and quality-review roles for independent cross-checking.
Human in the Loop
Requires approval or stop decisions before requirements and architecture baselines, safety decisions, document approval, dependencies, result acceptance, configuration baselines, releases, and assessor responses.
Guardrail
Restricts files and documents, commands, data, external communication, evidence-generation rules, merge, and deployment authority.
Traceability & Execution Evidence
Fail closed on missing or non-reproducible evidence
Narrating a plausible result is not the same as executing a tool.Preserve execution facts and stop rather than advancing on uncertainty.
Treat only results produced by actual approved commands and tools as evidence candidates
Preserve provenance, execution time, input and configuration versions, commit SHA, and tool versions
Inspect requirement-design-code-test-result links and report missing candidates
Never fabricate unexecuted test results, unperformed reviews, or nonexistent approvals
Fail closed when checks fail, provenance is unknown, versions differ, or approvals are missing
TDD, statement and branch coverage, static analysis, complexity, mutation tests, dependency audit, and latest-head CI are examples of customer risk-based internal gates, not A-SPICE requirements themselves
Adoption
Move from gap assessment to pilot, training, and measured expansion
Assess current processes and product risk, pilot one representative SWE flow, audit human review and real evidence, train development, quality, and safety roles, and expand only from measured results.
01 Scope, tailoring, and gap assessment
Assess product families, A-SPICE tailoring, ISO 26262 and SOTIF activities, current tools, security constraints, and accountability gaps.
02 Responsibility matrix
Classify AI-supported, human-only, and prohibited work plus approval, stop, and escalation ownership.
03 Agent, Skill, tool, and policy design
Connect Agents, Subagents, Skills, real tools, repository instructions, and evidence guardrails.
04 Representative SWE pilot
Run a small SWE.1-SWE.6 flow for document drafts, code, tests, and evidence.
05 Human review and audit
Review and revise drafts, execute code and tests, and audit traceability and evidence reproducibility.
06 Training and staged rollout
Train developers and quality staff and agree on the next scope from observed pilot results and limitations.
Deliverables
Leave drafts, review criteria, workflows, evidence rules, and limitations
Document deliverables remain review-ready drafts until accountable people revise, approve, and place them under configuration baselines.
Automotive, A-SPICE, and safety-activity AI adoption map
RACI or responsibility and approval matrix
Document draft templates and human review checklists
Agent, Subagent, and Skill design
Repository, data, tool, and evidence guardrails
Source and test code development workflow
Test automation and reproducible evidence pipeline
Requirements, design, code, test, and result traceability rules
Pilot results, limitations, and improvement priorities
Developer, quality, and safety-role training and rollout roadmap
Expected Outcomes
Reduce repeatable work and focus human attention on judgment
We do not guarantee compliance passage or numeric improvement.Goals are agreed from actual pilot evidence.
Less repetitive draft work
Prepare structured review-ready drafts while people retain approval and baselining.
More consistent code and tests
Connect approved requirements, design intent, and rules to Agent and Skill execution.
Repeatable verification
Re-run static checks and tests with the same commands, versions, and conditions.
Earlier traceability gap detection
Review missing requirement, design, code, test, and result links before baselining.
Reviewer focus
Automate organization and repeatable execution so experts focus on facts, safety, quality, exceptions, and approval.
Related A-SPICE consulting
Review the wider process assessment, definition, coaching, and pre-assessment service.
Related A-SPICE consultingRelated AI coding governance
Review the standard-neutral AI coding operating model, guardrails, and Human in the Loop.
Related AI coding governanceRelated AI CodeGen process
Review the code-generation and latest-head quality-gate process.
Related AI CodeGen processFAQ
Questions about automotive software, A-SPICE, and AI coding agents
Confirm safety responsibility, draft status, real execution evidence, and existing-tool integration before agent selection.
Does an AI coding agent guarantee A-SPICE or automotive safety compliance?
No. AI assists document drafts, code and test development, and actual verification execution, but it does not replace or guarantee A-SPICE compliance, assessment, certification, or ISO 26262 and SOTIF safety decisions. Process owners, engineers, quality and safety staff, assessors, and approvers retain judgment and accountability.
Does an AI-authored document immediately become an official deliverable?
No. It is a review-ready, high-quality draft, not a completed or approved deliverable. Accountable people must verify facts, consistency, traceability, safety, and project context, revise and approve it, and place it under configuration baselines before it becomes official.
Which activity should a pilot start with?
After assessing product risk and process gaps, start with a small flow with clear inputs and completion criteria, such as SWE.1 requirement drafts and traceability, SWE.3 code and unit tests, or SWE.4 real verification automation.
How are source, test quality, and real evidence checked?
Human code review and real build, test, and security checks are mandatory. Preserve provenance, execution time, input and configuration versions, commit SHA, and tool versions. TDD and coverage are risk-based internal gates, not direct A-SPICE requirements.
Can this work with existing tools and processes?
Yes. We first assess existing ALM, configuration, CI/CD and verification tools, internal AI platforms, Local LLMs, and security policy. Codex and Claude Code are examples only, and adoption remains vendor- and IDE-neutral.
Assess current processes and product risk, then start a small pilot connecting review-ready drafts, code, tests, real evidence, and accountable human approval.