← Back to Work
Case Study 02 — Internal Initiative

Bringing AI Fluency to a Design Org

How I lead the AI4Design working group at JPMorgan Chase — teaching designers to integrate LLMs and Copilot into their daily practice through workshops, onboarding documentation, and 1:1 ramp-up support.

Company JPMorgan Chase
Division Cross the org
Role Lead, AI4Design Workshops & Training
Timeline Ongoing 2026
Collaborators Designers across the org · non-technical to technical
01 — The Challenge

AI tools were available. Adoption was not.

LLMs and GitHub Copilot were rolling out across JPMorgan, and the promise for design work was real — faster ideation, structured research plans, working prototypes from markdown specs. But the tools were arriving faster than the practice around them. Designers had access to LLM, DevGPT and GitHub Copilot, but no shared answers to the basic questions: where does this fit in my day? what's it actually for? what does "good" use look like?

Most of the existing AI guidance was written by and for engineers. It assumed comfort with the terminal, with prompt engineering as a discipline, with code as a first language. Designers were being onboarded into a workflow that hadn't been designed for them.

Why This Was Hard

AI adoption in a design org isn't a tooling problem. It's a confidence problem. The barrier wasn't "can you use this?" — it was "is it for someone like me?" Without a designer-shaped onramp, most people would either bounce off or use the tools shallowly.

I saw this as a design problem in the most literal sense: the adoption experience itself needed to be designed — researched, documented, taught, and iterated on like any other product.

02 — My workflow

I built it on myself first

Before trying to teach anyone, I needed something worth teaching. I spent months integrating LLMs and Copilot into my own design practice — finding where they earned their place, where they didn't, and what the actual workflow looked like end-to-end.

[Image: Side-by-side comparison of payment flow variations across teams]
The result was a six-stage pipeline I now use on most projects. Each stage has a specific job, a specific failure mode, and a specific reason the tool helps.

The six stages

The result was a six-stage pipeline I now use on most projects. Each stage has a specific job, a specific failure mode, and a specific reason the tool helps.

PRD/DCE Assessment — Using LLMs to draft product requirements proactively when product partners request design without providing documentation — closing the gap so alignment can move forward without waiting on upstream deliverables.
Ideation & Variation Exploration — Brainstorming design directions, refining interactions, and generating variations faster than I could alone.
Research Planning & Documentation — Drafting structured research plans and variation details in markdown — streamlining the transition from concept to testable hypothesis.
Prototype Creation — Translating markdown specs directly into functional prototypes via Copilot, dramatically accelerating the path from concept to shareable artifact.
Collaboration & Sharing — Sharing working prototypes with product partners and engineers — enabling faster alignment and letting devs evaluate feasibility, unhappy paths, and error scenarios together during spike stories.

The biggest unlock wasn't any single stage. It was the continuity — having a defined workflow meant I stopped asking "should I use AI here?" and started asking "what does this stage need from it?"

03 — The Initiative

From personal practice to org-wide working group

Once the workflow stabilized for me, the question became: how do I make this transferable to designers who haven't yet found their footing with these tools? I founded the AI4Design working group as the answer with a few like-minded designers.

Designed as a product, not a memo

The working group is structured the way I'd structure any product with users on a learning curve: clear onramps for different starting points, low-friction support, and shared artifacts that improve as the community uses them.

1
Workshops I host directly
Live sessions walking designers through specific stages of the workflow — starting with the highest-leverage, lowest-risk entry points so people leave with a concrete win.
2
Workshops I moderate by recruiting peers
Identifying designers who've found a strong AI use case and giving them the platform — and the prep support — to teach it. This scales the program past me and builds internal AI champions.
3
In-person 1:1 ramp-up support
For designers who learn better by doing, I sit with them through their first real use case — debugging prompts, fixing tooling setup, getting them to a shipped artifact in one session.
4
Onboarding docs for non-technical users
Written specifically for designers, not engineers. No assumed CLI fluency, no jargon, no "and then just configure your env." The docs treat unfamiliar tools the way good design treats unfamiliar interfaces.
[Image: Workshop output — decision framework or Miro board]
AI4Design workshop: a live session walking designers through one stage of the workflow with hands-on prompts they leave with.

The breakthrough wasn't finding the "right" answer — it was structuring a conversation where twelve teams could discover the answer together.

04 — Enablement

Designing the docs the way I'd design a product

The onboarding documentation was the piece I spent the most time on. Most internal AI guides I'd seen failed designers in a specific way: they answered the wrong questions in the wrong order. So I rebuilt them from a designer's mental model first.

What's the first thing a designer reads?
Not setup. Not capabilities. A worked example — start to finish — of one real workflow stage with screenshots and the actual prompts I used. The reader sees the destination before the road.
How do you handle setup friction?
Setup is the highest-friction moment and the point where most non-technical users churn. I rewrote it with screenshots at every step, named every error message designers might hit, and added a "stuck? message Koni" line as the explicit fallback.
How is content organized?
By workflow stage, not by tool capability. A designer in research synthesis mode can find the synthesis page directly — not after reading three pages on what an LLM is.
How do the docs improve over time?
Each workshop and 1:1 surfaces a question the docs didn't answer. I treat the gap as a documentation bug — log it, fix it, ship the update. The docs get sharper every cycle.
[Image: Before — fragmented state]
Before: Payment method selection varied across teams
[Image: After — unified component]
After: Unified SafePay integration in the PDL

The Copilot disruption

Midway through this work, GitHub Copilot access was pulled across the org. The end-to-end workflow broke at a critical stage. The pipeline now has a well-defined input process (LLM-assisted ideation and documentation) and a hosting destination (Atlas), but the middle step —turning specs into interactive prototypes— is now manual and slow. The speed and iteration gains the tool provided are gone.

I documented the impact rigorously and surfaced it as a business case — both because the program needs it back, and because the loss is itself a teaching moment about how thin the seam between "AI-assisted" and "AI-dependent" actually is.

[Image: Key slide from the leadership presentation]
The deck focused on decisions made, not options explored — framing design work in terms leadership values.
05 — Outcome

What the program produced

6
Workflow stages defined and taught end-to-end
6
Hosted directly & moderated with peer instructors
+400
Onboarded on LLM, GitHub Copilot & Visual Studio

AI4Design gave the design org a shared language for AI-assisted practice — a pipeline, a vocabulary, a place to ask questions, and a set of designer-shaped instructions for getting started. The outcome I care about most isn't a number; it's that designers who previously avoided these tools now reach for them as part of their normal day.

The clearest signal of success: designers sharing their own AI-assisted work back into the community.

[Image: Slack/Teams thread or community board — designers sharing AI-assisted artifacts]
Slack/Teams thread or community board.
06 — Reflection

What I learned

On adoption as a design problem

The most useful frame I found was this: tool adoption inside an org is a product. It has users, a UX, an onboarding flow, and edge cases. Treating AI4Design that way — rather than as a training program or a memo — is what made it land. Designers recognized the shape of it because it was designed for them in the way they design for others.

On AI fluency as a design skill

I increasingly believe the line between "designer" and "designer who builds with AI" is going to dissolve the way the line between "designer" and "designer who works in Figma" did. AI4Design is my attempt to make that transition gentle — and to position the designers I work with on the right side of it.

Why This Matters for What's Next

AI is the most interesting thing happening in our craft, and the adoption question is wide open across the industry. The skills I built running AI4Design — designing onramps, writing for non-technical users, identifying what a tool earns its place doing — are the same skills that matter at companies building AI products from the inside. That's the work I want to do next.

Next case study
Building an AI-Era Product from Zero