Practical innovation for better science: How CRScube internalized AI to accelerate innovation

Introduction

This case study explores CRScube’s journey in adopting AI within its own software development lifecycle. By integrating Claude Code into our standard engineering and testing processes, we have transitioned from traditional coding methods to an AI-augmented approach. This shift was driven by a singular goal: to be more innovative ourselves so that we can deliver more innovative, high-quality solutions to our clients at an accelerated pace.

Background

At CRScube, our mission has always been to simplify clinical trial workflows. As the clinical landscape evolves, the demand for speed and precision has never been higher. To meet this challenge, we recognized that we could not simply add AI features to our products; we had to become an AI-driven organization from the inside out.

The primary objective of this transformation was to accelerate our development cycles while maintaining—and eventually enhancing—the rigorous quality and security standards required in clinical research.

The challenge: The context gap in LLMs

While Large Language Models (LLMs) are extremely powerful, they are strictly limited by the quality and depth of the input they receive. We realized early on that, without the full architectural context of an eClinical platform, even the most advanced AI produces generic or insufficient code. The challenge was not just finding a powerful model but learning how to provide the right environment for it to succeed.

An open-minded trial & error process

We approached this fundamental change with a "trial and error" mindset, accepting that failure was a necessary part of the learning process. We maintained a human-in-the-loop approach from day one, treating the AI as a new colleague rather than a black-box solution. We implemented strict peer code reviews, where our developers qualified every AI output just as they would for a human programmer.

The journey: From prompts to specifications

Our implementation followed a two-phase exploratory path:

• First attempt (prompting): Initially, we used traditional prompting to ask the AI to deliver specific segments of code. The prompts were elaborate and provided clear instructions, and expectations of the output. This phase failed; the AI lacked the deep context of our system architecture, resulting in quality that was lower than our internal standards.

• Second attempt (specifications): We re-evaluated our interaction with the LLM. Just as a human developer requires a detailed brief, we realized the AI required formal specifications. We moved away from simple prompts and developed a specification process designed specifically for AI.

• Success: This shift provided the AI with the necessary context. The results were successful, delivering high-quality outputs that met our standards, at which point we fully incorporated AI into our development processes. 

Re-skilling the team: A new way of working

Working with AI requires a different skill set than traditional team collaboration. Throughout the exploratory phase, our team was educated on these nuances. A prime example is our specification process: writing for AI is different than writing for a human. AI requires a level of explicit, structured detail that a human colleague might otherwise infer from experience.

We have reinforced to our team that the objective was never to replace human expertise, but to increase our output capacity. The AI handles the predictable tasks, allowing our developers to focus on high-level architecture, complex problem-solving and creative thinking.

Results: Practical innovation, proven impact

By successfully adapting our processes, we have achieved a fundamental transformation:

Want to see how our internal innovation translates to your study? Contact us to learn more about our AI-led EDC setup and automated testing.