AI in software development

AI assistants have long been an integral part of the daily operations of many software teams. Copilot suggests code, Large Language Models (LLMs) help with documentation, and test generators provide initial drafts. According to GitHub, four out of five developers use an AI assistant within their first week on the platform. In most companies, the question is no longer whether but how AI is used productively in software development.

At the same time, there is a growing sense that the hoped-for acceleration in development and testing is not happening, even though the tools are used daily. This article shows where the gap arises and how to close it.

The Cost of the automation gap

The consequences are gradual but measurable.

• The manual analysis effort required for pipeline errors, test maintenance, and incident handling does not scale with increasing project complexity. Recurring errors are not systematically tracked. Knowledge of their causes remains in the minds of individuals.
• Release cycles slow down even though more code is being written than ever before.

A gap is also becoming apparent from a business perspective. Research by McKinsey shows that only a small minority of companies derive measurable economic benefits from their AI investments. The DORA Report 2025 confirms this: Higher AI adoption does not automatically correlate with higher delivery performance. Without integration into the existing toolchain, new bottlenecks arise in testing, review, and deployment.

In regulated industries, such as medical technology and the automotive sector, another aspect comes into play. AI-generated artifacts require traceability and auditability, and without structured processes, documentation gaps emerge that will become apparent during the next audit.

From AI tool to AI workflow

The key lies in answering one question: How can AI results automatically flow into existing processes and trigger follow-up actions? Three workflow patterns have proven effective in practice.

1. Analysis ➜ Action
   Example: An LLM analysis of failed pipelines that directly generates a prioritized ticket
2. Generation ➜ Validation
   Example: AI-generated unit tests that are checked by an automated quality gate
3. Detection ➜ Escalation
   Example: AI classifies incoming incidents and routes them to the appropriate handling team.

A project from within our own ranks demonstrates how this works in practice. During an AI hackathon at a major medical technology company, the Pipeline Oracle tool was developed in just three days. It retrieves logs from CI/CD pipelines via API, analyzes them using an LLM, categorizes errors, suggests solutions with confidence scores, and automatically creates tickets on the project board. Reactive troubleshooting becomes proactive risk analysis.

However, Pipeline Oracle is a single use case and does not, on its own, answer the strategic question of where AI automation offers the greatest leverage. However, the underlying logic is transferable to unit test generation, incident routing, test framework migrations, and other processes.

Those who wish to systematically transition from experimentation to a productive workflow will benefit from a clear process model, defined entry criteria, and a plan for scaling and governance. This is exactly where our AI consulting comes in.

Sources

• GitHub: “Engineering leadership in the age of AI: Insights from GitHub / Enterprise Octoverse,” 2025.
• McKinsey & Company: “The State of AI,” 2025.
• Google/DORA: “Accelerate State of DevOps Report,” 2025.