AI-First Delivery: Running the Pilot the Right Way

We are running an AI-first delivery pilot inside our SCM platform — one of DSV's crown jewel systems — in close partnership with the DSV Digital Core team. The decision to move here, on a mission-critical vertical, was deliberate. If this operating model works at this scale and complexity, it works anywhere.

Guardrails before velocity.

The first investment was not in productivity tooling — it was in the harness around it. Coding agents operating at enterprise scale need a defined boundary: what they can touch, what patterns they must follow, how their output gets validated before it reaches a pipeline. We have put significant effort into AI guardrails that enforce our architectural standards, security posture, and compliance requirements. The agent gets autonomy within the fence, not outside it.

This is the part most pilots skip. They optimize for how fast the agent can generate code. We are optimizing for how safely that code can be trusted.

Where we are applying it.

The initial use cases were an obvious choice — high signal, lower blast radius:

• Test coverage and test suite modernization — generating tests for legacy code paths that never had them, and rationalizing existing suites that had grown brittle over time.
• New product coding — net-new feature work where the agent operates alongside engineers rather than on top of existing complexity.
• Incident analysis support — using AI to accelerate root cause analysis during live incidents, cutting the time between alert and understanding.

Measuring what matters.

The second area of investment is benchmarking. We are collecting structured data on productivity lift, not just developer sentiment. That means tracking where agents add clear value, and equally — where they introduce rework, hallucinate context, or create review overhead that erodes the gain. Pitfalls are data too.

Early signals are positive. The team is energized and asking to expand scope. We are taking the considered path: validate the current perimeter before widening it.

More to follow as the data matures.