If AI Can Build Products, What Are The Humans For?

AI can now write code, generate prototypes and ship functional products at a speed that would have seemed unrealistic even a year ago. But faster output does not answer the harder question: should this product exist at all? In his UXDX talk, Rory Madden, Founder of UXDX, explored where AI is already changing product development, where it still falls short, and why the future of human work may be less about producing outputs and more about owning outcomes.

The uncomfortable question behind AI product development

For a long time, the fear around AI in product development sounded distant. AI could help with snippets of code, generate ideas, produce copy, speed up research or create rough prototypes. Useful, certainly, but still very clearly a tool. That boundary is now harder to see.

As Rory Madden argued in his UXDX talk, AI is getting “scarily good” at building products. It can generate working code, create simple sites in minutes, assist with architecture, suggest UX patterns and reduce the effort involved in shipping software. For anyone working in product, design or engineering, the question is no longer whether AI will change the work. The more urgent question is what humans are still for.

That question can easily become abstract or alarmist, but Madden approached it more practically. He broke product development down into different types of products, different stages of the software delivery lifecycle and different risks. The result was a clear-eyed view of where AI is strong, where it is still weak, and how teams should think about their future role.

Simple products are already solved

Madden began by separating product development into three broad categories.

The first is low-interactivity products: marketing sites, brochure sites, blogs and simple one-way communication experiences. The second is the typical SaaS product, with forms, dashboards and moderate interactivity. The third is highly interactive software such as Figma, Miro, Google Docs or real-time collaboration tools. For the first category, Madden’s view was blunt: the problem is largely solved.

His example was his brother, who needed a simple brochure site for his company. After being frustrated by traditional outsourcing and weeks of poor results, he tried Lovable. Within 34 minutes, having never used the tool before, he had reached an output he was happy with.

That does not mean every AI-generated site will be beautifully crafted or strategically differentiated, but it does show the direction of travel. For simple digital products, the barrier to creation has collapsed.

Most product teams, however, are not only building brochure sites. They are building products with more interaction, more logic, more constraints and more business complexity. That is where the conversation becomes more interesting.

AI can code, but coding is not the whole job

When Madden turned to software delivery, he looked at the major stages: architecture, UI design, coding, QA, deployment and maintenance. QA and deployment, he noted, have already been heavily automated through years of DevOps progress. The more important question is what AI is doing to code, architecture, design and maintenance. On coding, Madden was clear: AI can write code, and it can often write better code than most humans.

He pointed to the release of Claude Code as a significant shift. What had once felt useful but limited became, in his words, better at programming than he was. It can now work for much longer stretches without constant correction, take on increasingly complex tasks and produce high-quality code with fewer prompts than before.

This matters because the coding bottleneck has changed. AI is no longer only helping with small tasks. It is starting to handle meaningful features, sometimes in one shot, with humans prompting, reviewing and correcting as needed.

But this is where Madden made an important distinction. Writing code is not the same as building the right product. AI can increasingly build the product right, but it is not yet reliable at deciding whether the product should be built.

Architecture is still where discipline matters

Architecture is harder than coding because it requires judgement about maintainability, security, scalability, performance and future change. A product can appear to work while being built on weak foundations.

Madden described AI’s architectural ability as inconsistent. Sometimes it produces strong decisions. Other times, a small change can lead it into strange or risky patterns. The issue is not that AI lacks knowledge in the abstract, but that it has access to enormous amounts of information. The issue is that it lacks the specific context of your business.

It does not know why a team made a decision last week. It does not automatically understand the trade-offs that sit inside the organisation. It does not have the accumulated, messy, human knowledge that often lives in people’s heads rather than in documentation.

That is why AI can become dangerous when used without expertise. Madden shared the example of someone who had “vibe coded” an app, posted about it online, and was hacked within minutes because the system had not been designed with security in mind. The app existed, but the architecture had not protected it.

This is the pattern product teams need to watch. AI can amplify the knowledge and discipline a person brings to it. An experienced developer can ask the right questions, spot the risks and guide the system away from poor decisions. Without that discipline, AI can produce something convincing but fragile.

AI still has a context problem

One of the reasons architecture remains difficult is that AI has a learning and context problem. Models know what was in their training data, but once released, they do not naturally keep learning your business in the way a team member does.

The industry is trying to solve this in several ways. Bigger context windows allow models to read more documentation and more code at once. Teams are creating files such as architecture.md or agents.md so AI systems can load organisational context at the start of a session. AI tools are adding memory features. Larger companies can fine-tune models on their own data.

But none of these solutions fully replaces lived organisational context. Documentation has to be written well. Context windows can fill up. Persistent memory is still limited. Fine-tuning can be expensive.

That is why human context remains critical. Teams know why decisions were made, which trade-offs matter and what the business is actually trying to achieve. Until AI can reliably hold and use that kind of context, humans remain essential to direction-setting.

Maintenance may be less of a problem than expected

A common argument against AI-generated code is that it will create huge maintenance problems later. If AI writes too much code too quickly, surely teams will suffer when they have to maintain it.

Madden challenged that assumption. Studies suggest AI can write cleaner code than many humans, especially when prompted with the right conventions and organisational patterns. It can also make maintenance easier by helping developers re-enter unfamiliar codebases, explain how systems work and identify where changes need to happen.

The key again is prompting and structure. If teams define how they organise files, separate responsibilities and document decisions, AI can support maintainability rather than undermine it. The risk is not AI-generated code by itself. The risk is AI-generated code without architectural discipline.

Design is behind, but improving quickly

UI design remains one of the weaker areas. Madden showed an example of a generated form that clearly needed improvement, and he described design as the biggest lag in the current AI product development stack.

But he also noted how quickly the space is moving. Tools such as Google Stitch, Figma Make, Claude design features, Lovable, Magic Patterns, UXPilot, Canva, Webflow and Replit are all investing in this problem. The quality is not yet consistently strong, but it is improving month by month.

This creates a challenge for designers. If AI-generated UI converges towards familiar, generic patterns, there is a risk of a bland sameness across products. But familiarity is also part of usability. Many products benefit from patterns users already understand.

The opportunity for designers may be less about manually producing every screen and more about knowing when to use familiar patterns, when to push for distinction and how to shape experiences around real user needs.

The real gap: building the right product

Madden used Marty Cagan’s four product risks to explain where AI helps and where humans are still needed: desirability, usability, feasibility and viability.

AI has dramatically improved feasibility. People who previously lacked the technical skills to build complex things can now do more. It also affects viability, because the cost of building software is falling.

But desirability and usability remain harder. AI can support research, summarise patterns and suggest improvements, but it does not automatically understand your specific customers in your specific context. It will not reliably tell you whether this feature matters, whether the problem is real or whether users will value the outcome. That is human work which becomes increasingly important as output becomes cheaper.

When it becomes easy to say, “we can build that in a week,” teams will be tempted to build too much. Without strong judgement, products can become Frankenstein systems: full of features, but lacking coherence, focus and real value.

Roles will change around outcomes

If AI shifts the cost of output, then teams will become more accountable for outcomes. Madden argued that the value of human work will move away from producing screens, tickets or code and towards deciding what should be produced and whether it worked.

That has implications for roles. Historically, teams were built around specialist skills: product managers, designers, researchers, front-end developers, back-end developers, scrum masters, analysts and more. The more people in a team, the more coordination was required, but organisations accepted that overhead because they needed the specialist skills.

AI changes that balance, it can absorb more tasks, which may lead to smaller teams or teams with broader scope. Madden pointed to the rise of design engineers as a sign of this shift: people who can both design and code, or at least move closer to production rather than stopping at static mockups.

He also predicted shorter iterations and more prototyping in production. If AI can generate production-quality code as quickly as a traditional prototype, then production itself becomes more disposable. Teams can probe, sense and respond with real users at greater speed.

Will AI cause job losses?

Madden did not avoid the question of job losses, but his answer was more nuanced than the headlines. AI can replace tasks, but it cannot yet replace full jobs. Product development roles contain many tasks, and AI may automate a large percentage of them, but the remaining work is still hard: judgement, context, alignment, strategy, accountability and understanding users.

He also pointed to Jevons Paradox: when something becomes cheaper, people tend to use more of it. As software becomes cheaper to build, more companies may be able to create products they previously could not afford. That could expand the amount of product development work overall, even as individual teams become smaller and more AI-enabled.

The work will not disappear, but expectations will change. People will be expected to do more, work across boundaries and take responsibility for outcomes rather than outputs.

What humans are for

Madden’s final answer was clear: AI can build products, but not necessarily the right products. That is where humans remain essential. Humans ask whether the product should exist. Humans understand the customer context, define the outcome that matters, set boundaries, make trade-offs, align stakeholders and decide when to stop building.

The future of product development may involve smaller teams, faster cycles and more AI-generated output, but that does not reduce the need for human judgement. It raises the bar for it.

Want to watch the full Talk?

You can find the full talk here: https://uxdx.com/session/if-ai-can-build-products-what-are-the-humans-for/

Or explore all the insights in the UXDX USA 2026 Post Show Report: https://uxdx.com/post-show-report