Sanitized Case Study - Accenture Ireland

Agentic AI SDLC Platform for Enterprise Prototyping

Building a governed Azure DevOps and Claude Code workflow for faster, auditable AI-assisted software delivery.

Azure DevOpsClaude Code Human in the LoopTraceability Security BoundariesEvaluation
ADOSystem of record
HITLReview and approvals
AuditTraceable artifacts

Context

The platform is designed for enterprise prototyping where AI-assisted implementation must remain inside delivery controls. Claude Code is an execution component; the operating model is the more important contribution.

Public-safe positioning: a Claude Code-integrated, model-adaptable AI SDLC pattern, with Azure DevOps retained as the enterprise control plane.

Workflow

Intake and requirements decomposition - solution design - implementation plan - AI-assisted execution - pull request - automated tests - human review - security review - release evidence - retrospective and evaluation.

Azure DevOps owns work items, acceptance criteria, repositories, pull requests, pipeline evidence, approvals, and release history.

Governance artifacts

  • Requirement traceability and design rationale.
  • Session summary, generated diff, and test evidence.
  • Risk flags, security considerations, review notes, and approval history.
  • Deployment-readiness checklist and known limitations.

Measurement layer

The evaluation design measures whether output met acceptance criteria, passed tests, followed repository conventions, introduced security concerns, or required significant human rework.

Useful operating metrics include cycle time from idea to pull request, acceptance rate, test-coverage delta, review rework, security findings, cost per prototype, and developer-reported time saved.

This approach connects delivery evidence to the public LLM Evaluation Workbench methodology.

Security boundaries

The public pattern emphasizes scoped repository access, secret exclusion, least-privilege tool permissions, human code review, retained evidence, approval gates, and escalation of uncertain outputs.

This case study intentionally excludes confidential client details, internal configuration, and unvalidated impact claims.