What 'AI product leader with design DNA' actually means
Design DNA is not a sensibility. It is a transfer function: the thing that turns a technically sound model into a product 400,000 people actually use.
Working draft, content pending
People use that phrase a lot. "Design DNA." "Design thinking." "Human-centered." It has been around long enough to become wallpaper.
So let me be specific about what it actually means. Not the conference-talk version. The version that shows up in the work.
I have spent 20 years doing this. Graphic design, UX research, product strategy, AI product leadership. I have shipped six AI systems inside a Fortune 5 company with 400,000 employees, and I led a procurement platform responsible for $20 billion in vendor spend. I have watched AI projects succeed and fail, and the reasons almost never have anything to do with the model.
This is what "design DNA" actually means when you are building AI products at scale.
Jensen Huang's stack and mine
Jensen Huang describes AI infrastructure as a stack. Physical infrastructure at the base, then data, then training, then inference, then applications at the top. Each layer enables the next. You cannot skip a level.
I think about human capability the same way. Over the past two decades I built a stack too. Not in servers. In skills and ways of thinking.
Layer one: visual language. BFA in graphic design. Years of print, web, digital. Understanding how humans read a page, how color creates hierarchy, how whitespace is not empty, it is structure. You learn to see systems in a layout before you can explain why.
Layer two: user research. Hundreds of user interviews. Sessions in the US, UK, Canada, Germany, Australia. Usability tests, design thinking workshops, shadowing people at their actual jobs. You learn that what users say and what they do are different things. You learn to find the real problem underneath the stated problem.
Layer three: systems thinking. Design systems, component libraries, governance models. How you encode decisions so they scale. How constraints enable quality instead of restricting it. How you make things consistent across teams and time.
Layer four: product strategy. Connecting design and research to business outcomes. P&L responsibility. Stakeholder management across functions that do not agree on anything. Translating between the language of users and the language of executives.
Layer five: AI product leadership. Taking everything above and applying it to systems that learn, predict, generate. Building for probabilistic outputs. Designing for uncertainty. Shipping AI that employees at a Fortune 5 company actually use and trust.
Each layer depends on the ones below it. You cannot skip a level.
The sixth layer
Here is the part Jensen's stack does not have.
There is a sixth layer in human capability, and it is the only one that cannot be automated. Call it the human layer. The capacity to sit across from another person, understand what they are actually afraid of, and design something that addresses that fear while still meeting the business objective.
Empathy is not a soft skill. It is infrastructure.
I will show you what I mean.
In 2022 we had a problem. Enterprise procurement at UnitedHealth Group, the platform that now manages $20 billion in vendor spend and more than 8,500 contracts a year, was running intake on a 15-minute manual questionnaire with a 900-category taxonomy. Business requesters picked the correct category 40% of the time. Which means 60% of requests were routed to the wrong place, bounced back, needed rework, caused delays, and frustrated everyone.
The technical solution was obvious. Train a classifier. Automate the categorization. Problem solved.
Except that is not the problem.
The real problem was that requesters did not understand the taxonomy. They did not know what "indirect spend, subcategory: software and SaaS, volume: enterprise license" meant. They were not procurement specialists. They were a benefits coordinator trying to get a software renewal approved. A facilities manager trying to hire a vendor. People using a system designed for procurement experts, when the actual users were just people trying to get things done.
The solution was not a better classifier. The solution was replacing the form entirely with a conversational AI that met users where they were. Ask them what they need in plain language. Let the AI figure out the category. Accuracy went from 40% to 85%, and the savings were $390,000 in the first year.
That is empathy as infrastructure. Understanding who is actually using the thing, and designing for that person. Not the ideal user. Not the technical spec. The actual human who is stressed and busy and just trying to get through their day.
The engineering team could have built a better classifier. Only someone who spent years doing user research would have reframed the problem entirely.
Design DNA as a transfer function
In signal processing, a transfer function describes how a system transforms an input into an output. The input goes in. Something happens inside the system. A different output comes out.
Design DNA works the same way.
The input is a technical AI capability. A language model, a classifier, a recommendation engine. The transfer function is everything I described above: the visual language, the user research, the systems thinking, the empathy. The output is a product that humans actually use.
Most AI projects die not because the model failed, but because nobody applied a transfer function. The model works in the test environment. It does not work for the 400,000 people who need to use it in their actual jobs.
Here is a concrete example. We built an autonomous customer service resolution system. AI monitors all incoming requests across a CRM. When a case has a known answer at 80% or higher confidence, the system generates a response, sends it to the customer, and closes the case. No human involved.
The engineering spec is one paragraph.
The product design is months of work.
What does the customer experience when they receive an AI-generated response? Do they know it was AI? Do they care? What happens when the AI is wrong? Who is responsible? How does the escalation path work? What does the handoff to a human agent feel like? Does it feel like a drop, or does it feel like care?
We went from 100,000 human-handled emails per year down to a projected 60,000. That is 40,000 cases resolved autonomously. Those numbers are real. But getting there required designing every edge case, every failure mode, every customer touchpoint with the same rigor you would apply to a consumer product.
That is the transfer function in action.
The model does not know how to do that. The product leader with design DNA does.
The hardest users in the world
There is a version of AI product work where adoption is free. Consumer apps get curiosity. Startups get early adopters who want the product to win. Demos get audiences who came to be impressed.
Enterprise employees are the opposite audience. Nobody at a Fortune 5 company wakes up wanting your AI tool. They have a job, a backlog, and a decade of memories of software that promised to help and added work instead. They do not extend the benefit of the doubt. They try the thing once, and if it wastes ten minutes, they route around it forever, and they tell their whole team to do the same.
That is the audience I build for. Skeptical, busy, and free to ignore me.
Which is why I trust enterprise adoption as a signal more than almost any other metric in AI right now. Benchmarks measure the model. Downloads measure curiosity. But a claims processor who chooses the AI path on a Tuesday afternoon because it is genuinely faster than the old way? That is a verdict on the whole product: the model, the interface, the failure handling, the trust.
If you can make employees at a company of 400,000 love an AI product, you can make anyone love an AI product.
Twenty years is not a liability
I keep hearing the same narrative. AI moves fast. Traditional skills are being disrupted. You need to be technical to stay relevant. The implication, sometimes explicit, is that twenty years of design and product experience is a drag on your ability to operate in this world.
I think the opposite is true.
Not because the experience is valuable in some abstract, nostalgic way. Because of what it specifically teaches you that you cannot learn any other way.
You learn to be wrong. Designers get feedback constantly. Users ignore your work. Executives kill your projects. Stakeholders reject your recommendations. You internalize that the output is not about you. It is about whether it works for the person using it. That takes years to actually learn, not just say.
You learn that the first solution is usually not the right one. Research reveals that. Testing reveals that. Iteration reveals that. AI makes it easy to generate a first solution fast. It does not make first solutions better.
You learn that systems degrade. A design system without governance becomes inconsistent within 18 months. A product without a feedback loop drifts from user needs. An AI system without ongoing evaluation starts making decisions nobody intended. Managing the long arc of a product is different from shipping it.
You learn that trust is slow to build and fast to lose. Every user interaction either deposits or withdraws from a trust account. AI products that are inconsistent, opaque, or wrong at the wrong moments will be abandoned, no matter how impressive the underlying model.
Twenty years of that is not a disadvantage. It is pattern recognition that no training set has yet captured.
What the phrase actually means
Let me make this concrete.
It means I can read a technical spec and a user journey map with equal fluency. I can sit in a machine learning review and ask the right questions about training data and evaluation criteria. I can sit in a design critique and know immediately when a flow is designed for the system's convenience rather than the user's.
It means I think about AI outputs as experiences, not just outputs. The confidence score is not just a number. It is a signal that needs to be communicated to a human in a way they can act on. The generated response is not just text. It is a message from an organization to a person, and it carries all the weight that implies.
It means I default to asking "who is this for?" before "how does it work?" That question sounds simple. In practice, in meetings with engineers and executives and stakeholders who all have different answers, it is the hardest question to keep asking.
It means I understand constraint as design. The best design systems are not collections of unlimited options. They are sets of decisions already made, boundaries that enable speed and consistency. The best AI systems are not maximally flexible. They are governed. They know what they will not do as clearly as they know what they will. That is not a limitation. That is a feature.
It means I have done the unglamorous work: the 500 user interviews, the usability tests, the workshops, the failed prototypes, the products that shipped and still did not work for the people they were built for. That failure teaches you things that success does not.
The honest part
I am still learning the technical layer. Actively, and on purpose. I am deep in advanced AI training and certification work, and I am pushing to understand the underlying systems, not just the outputs.
I know what I do not know. I know where I need engineers and data scientists. I know how to work alongside that expertise without getting in its way and without letting it run unchecked.
That is the collaboration. Not "designers versus engineers" or "business versus technical." The design DNA is the translation layer. It is what turns a technically sound model into something a real person will trust and use.
I have built systems for 400,000 people. I have watched those systems succeed and fail. I have learned more from the failures.
The phrase "AI product leader with design DNA" is shorthand for a specific set of experiences and capabilities. It is not a claim about having all the answers. It is a claim about having a way of working that produces better AI products.
Not better in the benchmark sense. Better in the "does this actually work for the people who have to use it" sense.
That is the only measure that matters.
One more thing about stacks
Jensen's stack and mine share a property worth naming. The deepest layers are where the leverage lives. Applications come and go at the top; infrastructure decides what is possible.
I have spent twenty years working from the top of my stack downward. Lately I have been descending it on purpose, closer to the systems themselves, closer to the layer where AI products are actually built.
Why I am doing that deliberately, and what I think it means for where product leadership is heading, is the next essay.