Summary
Key results
80–88% less time on repeatable integration tasks
Up to 84% generated code assessed as correct on the first attempt
When growing demand outpaced engineering capacity
Sii Poland’s Internal IT teams develop and maintain a broad technology landscape in parallel, including external web and mobile applications, integration services, enterprise resource planning, business process management, a Microsoft Dynamics 365-based resource management system, and the Azure DevOps delivery pipeline. The estate includes more than 115 applications and platforms and over 280 integration modules connecting more than 30 internal and external systems.
As demand for new functionality increased, the breadth of this landscape put growing pressure on available engineering capacity. Several recurring constraints absorbed time that could otherwise go into higher-value delivery:
Repetitive engineering work
New front-end components often required the same files, tests, translations, and documentation, while new integration branches could require hours of manual scripting.
Code review load
A medium-sized change could take 45–60 minutes of a senior engineer’s time, and developers could lose momentum while waiting for review.
Concentrated system knowledge
Consultants and testers often depended on a small group of senior specialists for answers about system behavior, which could delay work by hours or days.
Low-value improvements left in the backlog
For some small automations and quality-of-life enhancements, traditional implementation effort exceeded the expected business value.
Further actions needed
Early experiments with generic artificial intelligence assistance did not solve the problem. Without project architecture, conventions, and dependency context, generated code often required rework, while developers spent time repeatedly explaining the project before the model could contribute effectively.
Sii Poland therefore assigned the initiative to its own internal engineering teams – specialists who also deliver software engineering and artificial intelligence adoption services to commercial clients.
Context-rich developer kits standardized Claude Code use
Sii Poland’s engineering teams did not roll out Claude Code as a generic assistant. For each of 9 technology areas – from enterprise resource planning to mobile development – they built a developer kit. This kid gives the model the same core context a new team member would need to work within that team’s standards.
The kits combine a common set of building blocks:
- Project rules – team-specific architecture, coding conventions, and standards that keep generated work within defined patterns
- Live project knowledge – Model Context Protocol servers that expose component indexes and a regularly refreshed view of dependencies across database, process, and integration layers
- Specialized agents – dedicated assistants for implementation, code review, test automation, answering questions about system behavior, and turning logs and screenshots into Jira-ready bug reports
- Repeatable commands – single commands for component scaffolding, code reviews, financial dimension setup, deployment checks, and log health checks
- Built-in safeguards – automatic blocking of files containing secrets, test execution after code generation, and mandatory developer review before changes enter the codebase
- Automated pull-request review – an artificial intelligence-assisted Azure DevOps extension that comments on pull requests and checks 20 security rule classes before code reaches production
Sii Poland also standardized distribution and maintenance. Each developer kit installs with a single command, is versioned together with the relevant codebase, and can be improved by developers as part of daily work.
Lower delivery effort made more improvements viable
Lower implementation effort changed which work made business sense. Teams could bring long-postponed automations and quality-of-life improvements back into sprint planning because they no longer required the same amount of engineering time to deliver.
Delivery also became more consistent. Code reviews follow defined checks, bug reports use a standardized structure, and generated code reflects team conventions earlier in the process. Project knowledge encoded in rules, agents, and knowledge servers gives consultants and testers another route to answers, reducing dependence on a small group of senior specialists.
Because the same pattern worked across 9 technology areas with different technology stacks, Sii Poland also gained a repeatable internal model for context-rich Claude Code adoption. Experience from operating it on Sii’s own systems can now support similar artificial intelligence adoption initiatives for commercial clients.
Measured gains across delivery, review, and reporting
- Up to 30% faster delivery on selected internal projects
- First-pass correctness of Claude Code-generated code increased from about 33% to about 84%
- 80–88% less time on repeatable integration tasks – financial dimension setup fell from 20–24 hours to 3–4 hours, and unit test creation from 12–16 hours to about 2 hours
- Code review time reduced by 40–55%, with estimated savings of about PLN 60 per pull request and PLN 156 000 per year at 50 changes a week
- Bug-report preparation cut from up to 15 minutes to under 2 minutes, while report completeness increased from about 60% to about 95%
- About 50% less effort required to deliver Azure DevOps extensions