Data scientists often struggle to deliver impact in production because they work in notebooks, outside the application stack. Could GenAI tools for UI creation help solve the low ROI problem in data science?

Reconsidering My Position on No-Code

I recently came across an AI tool for designing web applications. A client of mine was eager to try it, but I had some reservations. Almost exactly 4 years ago I said publicly that I’m not a fan of low-code/no-code tools.

While UI-generator tools (I’m including the old-fashioned drag-and-drop ones here as well as the new AI-powered ones) are neat, my instinct is to worry about things like long-term maintainability and customization, especially for non-trivial applications and tricky data models. Imagine trying to untangle a big Python or React project built with a drag-and-drop interface designed for non-technical users. Maintaining a codebase generated by a no-code tool would be like inheriting a project from a prolific coder who somehow never developed preferences or a discernable coding style.

But it’s important to keep an open mind. A lot has changed in the last four years and I’ve come to integrate several AI tools into my everyday workflow. Perhaps the lady doth protest too much.

So I asked a group of my software engineering buddies what they thought, and it totally changed my perspective. They said AI-generated code may not be inherently worse to maintain than code written by humans who’ve had to prioritize speed over long-term maintainability. 🤯

After all, these AI for UI tools are appealing for good reason; things are certainly a bit off in the state of DenmarkJS…

The Case for a Disposable UI Mindset

Perhaps the problem is that we are clinging to the idea of a perpetually maintainable and perfectly optimized user interface.

Think about it: how many times have you inherited a codebase from outsourced contractors or even well-meaning colleagues from different backgrounds? These projects often end up needing a rewrite anyway, even if they are framed as maintenance efforts at first.

There are a lot of reasons for this. For one thing, the JavaScript world moves at breakneck speed. New frameworks and tools emerge constantly as others become obsolete. This makes it incredibly difficult to build UIs that remain relevant and maintainable over the long term. There’s also a lot of turnover in our industry, especially recently with all the tech layoffs. This means we are all constantly inheriting code from predecessors with varying skill levels, coding styles, and understanding of the project. Inheriting code, whether from skilled developers, less skilled developers, or an AI, often comes with a significant amount of tech debt. From suboptimal design choices, to undocumented code, to dependence on outdated or unsupported technologies, tech debt only ever seems to accumulate.

With all of that in mind, taking a “disposable UI” mindset is oddly liberating.

Thinking in Applications

But maybe we should be thinking bigger. Might “disposable UI” also help solve data science’s low ROI problem?

Data scientists excel at exploring data in notebooks, but their larger-scale efforts often fail because building a successful data product requires a mental model of the entire application stack. This is why I coach data teams to shift towards “thinking in applications” (i.e. as opposed to thinking in Jupyter notebooks or PowerPoints). For instance, how will the model be deployed and accessed? How will other applications interact with the model? How will users interact with the model’s outputs and gain insights?

Learning to ask these questions is how data teams deliver truly impactful and user-centric data products. But it’s a lot to ask of people to learn both data science and applications development. And this is the use case I have in mind for AI-generated UI tools.

In the hands of data scientists, AI-powered tools that generate UI prototypes based on simple descriptions or sketches (Figma, Adobe XD, Uizard) might make it easier to think at the scale of the application. By reducing friction in UI development, I could see how these AI tools might teach data scientists to prioritize user experience and to reason about the interactions users will have with their models. And if the first UI attempt sucks or needs significant updates, the “disposable UI” mindset makes it easier to fail and learn.

Nothing good or bad, but thinking makes it so

While AI tools can generate functional UIs, do consider the long-term maintainability of these interfaces. My front-end buddies made it clear that they wouldn’t recommend using these tools for more complex applications. But for rapid prototyping or similar greenfield projects, genAI tools for UI creation might help empower data teams to “think in applications” and more successfully deliver impact in production.

Tried one of these tools and have an experience to share? I’d love to hear from you!

AI Image Generated with Imagen 3