From fragmented R&D to an AI-enabled delivery system

The operating challenge

The context was one of the country’s largest private, multinational businesses: an organisation employing 55,000 people, serving 5.6M+ household and business customers and running substantial internal R&D. Many product streams moved in parallel, while the business around them kept changing.

The pressure was not simply to make a new interface faster. Every change could create a support burden in work already delivered: adjustments, improvements and refactoring competed for the same people and capacity as new initiatives. The question was how to increase R&D throughput without losing control of consistency, quality or business accountability.

My mandate: make AI operational, not ornamental

I was appointed to lead development–design synchronisation and to investigate how AI could improve the operating model. I worked alongside the development team to define an AI-first workflow architecture: one that gives AI useful, governed context instead of asking it to invent a design or implementation from a blank prompt.

The principle was simple: maintain the foundations once, in places the human team can realistically keep healthy, then let those foundations guide design and implementation. That reduces unnecessary rediscovery and makes changes easier to trace back to a shared standard.

The AI-first architecture

The UX pattern library in Figma became the human-maintained source of truth for experience patterns, states and interaction decisions. With development, we connected it to a production frontend component library, including Figma Code Connect, so design intent and coded components could be understood as parts of the same system rather than two separate handoffs.

On top of those two foundations, I created an agent and a set of task-specific skills. MCP-connected tools and APIs gave the agent controlled access to the relevant systems; retrieval-augmented generation (RAG) and vector memory supplied the right pattern, component and project context for a request. In practical terms, the AI could work from the team’s established knowledge instead of relying on generic assumptions.

From manual design to prompt-first design

The workflow changed from designing every screen manually to a prompt-first, AI-enabled process. A designer frames the need and prompts the AI. The AI consults the Figma UX library and the production frontend library, then proposes an interface solution that starts from the available patterns and components.

The designer remains the evaluator. They inspect the proposal for errors, mismatches, missing states and poor judgment; simple corrections can be made by prompting, while nuanced work is refined by hand. This is not “design without designers.” It is a faster first pass with a qualified human responsible for the final design judgment.

Human approval stays in the system

Once the design is ready, it moves through business and stakeholder approval. That gate is deliberate: AI can accelerate exploration and implementation, but it does not decide what the organisation should release. Business owners remain accountable for the priority, scope and final direction.

Only after approval does the AI take the final design into implementation. It uses the same frontend library and the same UX source of truth to generate the product work, while the design and development teams review, correct and extend it where the problem requires human expertise.

What this changed

The result is a repeatable, human-governed design-to-code operating model: shared standards are maintained once; AI can retrieve them when creating; designers and stakeholders make the decisions that require judgment; and implementation begins with less distance from the approved solution. It is designed to make a dynamic R&D environment more controllable, not to remove the people who make it work.