Building design practice from scratch at an Enterprise AI platform, then rebuilding how design works when AI changed engineering

Kore.ai builds the platform Fortune 2000 companies use to create and deploy AI agents. When I joined as Director of Product Design in August 2023, the platform was technically sophisticated and the design practice had been ignored. Multiple teams were building in isolation, each with their own patterns. No shared design system, no tokens, no design review process. Four products that did not feel like one platform. Three years. Three chapters.

Chapter 01: Unify

The fragmentation had real costs. Inconsistent patterns meant engineering teams were rebuilding the same components across products. Users had to relearn basic interactions when moving between parts of the platform. There was no shared language between design and engineering because there was no shared system.

The question was whether to start with a component library or with tokens. A component library is the more visible deliverable. Tokens are the infrastructure that makes a library consistent. Starting with components and adding tokens later means retrofitting, which rarely holds.

I chose tokens first. That meant the first months produced nothing the organisation could see. The tradeoff was worth it: once the token layer existed, components built on top of it stayed consistent without requiring manual enforcement after every update.

The result: a coded design system with live tokens across all four products, a UI component gallery documenting every shared pattern, and a design review process that had not previously existed.

Chapter 02: Build new

The company built a multi-agent agentic automation product from scratch. No established design patterns existed in the industry for this. Enterprise teams needed to configure, test, and deploy autonomous AI agents in banking, healthcare, and customer support.

The design challenge was not the configuration UI. It was trust. AI systems can generate different outputs for the same input. Enterprise customers deploying AI in regulated industries cannot accept unpredictable behaviour as a known condition. They need confidence the system will behave as expected.

Competitors were either hiding AI behaviour behind abstraction, which enterprise customers rejected, or surfacing it with full technical detail, which non-technical administrators could not use. Neither was viable.

I designed for graduated transparency: a default view showing agent outcomes and the key decision points that shaped them, with drill-down access to underlying reasoning for users who needed it. Users could intervene at defined checkpoints rather than monitoring every step. The tradeoff was that the default view required careful curation. Every iteration involved deciding what to hide as much as what to show.

Chapter 03: Transform how design works

The product shifted to vibe coding. Engineers were generating UI directly from AI prompts. The design review process built in Chapter 1 no longer fit the cadence. What had worked for a weekly review cycle was now a bottleneck.

The question was whether to maintain Figma-based workflows or redesign how design work happened to fit the new reality. Maintaining Figma-based workflows would have preserved familiar process but made design increasingly irrelevant to how the product was actually being built.

I researched how 11 companies were handling design within AI-assisted development workflows and proposed a four-layer integration model to the SVP and CTO: automated enforcement with design standards embedded in AI coding tools and CI linting; on-demand pairing where design is available when engineers need it; weekly audits where the design team scores shipped product and fixes directly where possible; and design-led initiatives for user research and cross-product coherence.

The team moved to a Figma-less workflow, designing directly in Claude. I built five Claude-powered tools: UX Orchestrator, UX Audit, Product Experience Audit, Design Review, and Design Review Handoff. Delivery timeline moved from weeks to days after the workflow shift.

What I would do differently

The audit baseline of 51/100 across 55 pages was diagnostic. I set a score but not a timeline with defined milestones. That made the audit a reporting tool rather than a change management tool. Running this again, I would establish quarterly targets tied to specific pages or product areas rather than a single aggregate score. That way the team can see directional progress instead of a number that moves slowly and is hard to act on.