Ben Woodmansee — Designing Frontier AI Products

View Case Study

Hey, I'm Ben. Design Lead for AI Incubation at IBM, where I turn frontier AI research into products people actually use. Before this, I led design across hundreds of generative AI projects at IBM Client Engineering (the ones I'm proudest of: NatWest's Cora+ and a £400M UK Government programme), and before that I was Founding Product Designer at Grandeo Labs, a ML startup later acquired. Erna Plachte Scholar at University of Oxford. Forthcoming Fellow at the Oxford AI Safety Initiative (OAISI). Studying AGI Strategy at BlueDot Impact.

I’m a Design Lead with 8+ years experience building frontier AI products, from computer-using agents and agentic systems to vibe-coding platforms and generative tools used by millions.

RECENT PROJECTS

(02)

Selected

↓

Work

Projects from years
©2017 — 2026

A few recent projects exploring how new AI capabilities reshape the interfaces we build.

AI Incubation

Design Lead at IBM’s AI Incubation lab building early-stage AI products. The team operates as a 0 → 1 environment focused on agentic systems, developer tools, and new interaction paradigms.

0 - 1 AI Product Design

AI Prototyping

Hands-on Building

Project Flint - Vibe Coding Platform, 2025 - 26

Led the 0→1 design of an internal natural-language-to-application builder. Shipped MVP in 4 weeks and scaled to 500+ users in 12 weeks by redesigning the interaction model around iterative building.

Design Lead

0 → 1

Vibe Coding

Automating Correspondence with Generative AI, 2024

Designed an AI system for a £400M government transformation programme that automates large-scale correspondence workflows while maintaining auditability and policy compliance.

AI Governance

Generative AI

Government Scale

NatWest's Cora+ with Generative AI, 2023

Led design of NatWest’s Cora+ generative AI assistant used by 11M+ customers. Built guardrails and interaction patterns to ensure the system could never hallucinate financial advice.

Generative AI

Assistants

Consumer Facing

Dynamic Quantum Optimisation - IBM Quantum, 2024

Designed a quantum optimisation interface translating complex quantum simulations into decision workflows usable by financial analysts.

Whitelabel

Prototyping

Quantum Computing UI

Other Product Design Work

See more of my work: AI-assisted conversation design for a FTSE 100 client, UCL's first unified design system (BIMA Award Winner), and a patented blockchain file-sharing platform acquired by Binarii Labs.

View all projects

(03)

What I Work on

I work across the full stack of AI product design and development: from prototyping with code and designing agent supervision systems to shipping generative tools at enterprise scale. These are the problems I find most exciting, and the projects that shaped how I think about them.

→
0-to-1 Frontier AI Products
I enjoy working on products before there's a clear playbook for what they are. Flint and Relay are the clearest examples. We shipped an MVP in four weeks and scaled to 500+ users in twelve. The early product was framed as a code generator but that was wrong. What users wanted was not perfect code, they wanted an accelerated way to prototype ideas, see them take shape, and then iterate. The product started working once we shifted from one-shot prompting to fast draft, edit, and repeat.
→
Prototyping with Models
AI products now move too quickly for long design cycles. I prototype close to the real system, often in code, so the interaction patterns can be tested while the product is still taking shape. On Flint, that meant shipping continuously and testing weekly. On Webglide, it meant working directly with the agent framework to understand how control should actually work. With AI systems, the real design problems only show up once the model is in the loop. This increases your ability to ship fast, collect feedback and iterate.
→
AI Safety & Governance
Some of the most interesting design problems in AI come from constraints and governance. On NatWest’s Cora+, the system could never hallucinate financial advice to different personas. On the UK Government programme, generated output carried institutional and legal weight. On Flint, trust was dependent on Carbon alignment, reliability, and visible failure states. That changed how I think about AI design. The challenge is often not making the system more capable, it's defining where autonomy is dependable, where review is needed, and how those boundaries become legible.
→
Systems at Scale
I like early-stage work, but I also care about what happens when systems meet real complexity. Cora+ served 11 million customers and was one of the largest genAI deployments in UK at the time. The UK Government work sat inside a £400M transformation programme. Flint scaled through continuous internal releases and was often used in client-facing contexts. At that point, you are not designing for ideal users in clean flows. You are designing for legacy systems, compliance, uneven trust, and real organisational mess. That is where the problem space gets interesting.
→
Tools for Developers & AI Infrastructure
Neurolab started because we observed how developers struggle to debug failing agent workflows with no visibility into what went wrong. Our team designed a tracing and observability tool for LLM applications that made multi-agent systems, tool calls, and self-healing loops legible to the people building them. I've also built internal tooling like a Figma AI Classification asset used by 400+ IBM designers. I'm most interested in the tools that help other people build: the infrastructure layer where good design has an outsized impact.
→
New Interaction Paradigms
I design interfaces for systems that can generate, reason, and act. Most of my work sits between product design, research, and hands-on building. I’m passionate about the moments when a new model capability appears, and the old interface stops making sense, and invention has to take place. We're moving to a homegenised but custom product world, where interfaces are adaptable, steering but importantly are used to steer, supervise, and build with AI, not just prompt it.
→
Agentic Systems & Agents
I designed the interaction layer for a computer-using agent built on IBM’s CUGA framework (#1 on AppWorld, #2 on WebArena). The core design question was simple. If an agent can operate any interface, what does the human need to see? Supervision interfaces change what people feel comfortable delegating. Is a webpage that useful for a user to see? How do you present users with what they actually value?

CREATIVE RESEARCH & TOOLS

(04)

Creative Research & Tools

Outside product work I’m a writer, artist, and electronic music producer. I write essays on AI history and design futures on World Make, my Substack. My fine art training taught me to treat research as a creative practice and as leisure.
‍
On weekends I’m as likely to be rebuilding Gemmascope as patching a Moog Modular.

Classification with Generative AI - Figma

Created an AI classification asset used by 400+ IBM designers, helping standardise taxonomy and accelerate design system adoption across the organisation.

Neurolab, 2025

A tracing and observability tool for LLM applications designed for enterprise developers. Helps debug complex agent workflows including multi-agent systems, tool calls, and recovery loops.

Editor - Next to Nothing - Topsafe London

Edited a casebook for filmmakers, artists, and image-makers exploring independent filmmaking through stories and interviews. Featured in i-D Magazine, It's Nice That, topsafe.co.

(05)

Awards

2016

Erna Plachte Scholarship

University of Oxford

→

2018

Next-to-Nothing - iD Magazine

Topsafe

→

2021

The Drum Awards - 4 Nominations

Awarded UK Government Technology Grant

Grandeo Labs

→

2023

Sustainability Partnership of Year

IBM

→

2024

IBM Technology Excellence

Oxford AI Safety Initiative

Follow Me On Substack
for weekly essays on Frontier AI and Design Futures

Explore Page

→