The increasingly valuable skill of prototyping with AI

It is well-documented that, as the AI community delivers better models, the end-user cost of running these models decreases over time. When a new algorithm is introduced, incremental innovations in research, better implementations, and new generations of hardware contribute to a “multiplier factor” in efficiency after some time. An OpenAI paper of 2019 estimated that, with state-of-the art implementations and up-to-date hardware, the cost in resources of training an image classifier as good as AlexNet, a breakthrough in 2012, had been cut by 44x in 7 years. 

This cost reduction couples with another “ecosystem phenomenon”: AI is not only more efficient, but also more commoditized. Compared to, say, 15 years ago, the delivery of AI has converged to a few “dominant designs”: popular open-source libraries, easy-to-access software APIs or model hubs, facilitating the integration with other pieces of software. Classifying an image with a deep learning model like AlexNet would have been a non-trivial effort for a competent programmer in 2012. Today, the model can be downloaded from a platform like Huggingface and queried with a wrapper library—and the same holds for a vast collection of models, including some of the latest large language models (LLMs) like DeepSeek. Therefore, AI is not only cheaper in terms of necessary resources, but also cheaper in terms of breaking the barrier of adoption. 

One of the consequences of this trend has been postulated by some economists: as the cost of AI predictions goes down, the value of skills that are complementary to AI will increase.  

In this text I argue that one such complementary skill that will become more and more valuable is prototyping with data and AI. Until recently, prototyping a digital product was viewed as a costly and specialized activity. With the current ecosystem of commoditized tools for AI and application development, and with the assistance of generative AI, we need to redefine that vision.  

From traditional to AI prototyping

Prototyping refers to the process of rapidly generating and testing product ideas by creating a “low fidelity” representation of the final product, deliberately unpolished in terms of appearance or functionality. This process serves as an effective tool for reflecting by observation and gathering feedback from stakeholders, prior to investing significant resources in more costly phases of product development.  

Prototyping techniques try to answer the question “are we about to build the right thing?”, which traditionally has translated to answering questions about the appearance or user experience, through techniques including paper prototypes, wizard-of-oz prototypes, wireframes, mockups or design fictions.  

In contrast, when it comes to data and AI products, prototypes should specifically ask two additional questions:  

1. Is the technical hypothesis involving data and AI viable?  
2. Is the accuracy of the system good enough?  

It is almost inconceivable to assess the quality of data, the predictive power of variables, or whether the final accuracy of the system delivers the desired experience without constructing a software prototype and iterating for real.  

Prototyping with AI demands a discovery-driven mindset and an intuitive sense for the business needs

Until a few years ago, building such a software prototype was viewed as transitioning to a more advanced phase of fidelity with increased expense. In the last years, technical roles such as data scientists, business analysts and AI engineers have witnessed how the software development ecosystem has been populated progressively with more and better tools, perhaps reaching a “tipping point” that brings to the next level what previously was considered prototyping. Now a developer working with the Python ecosystem can exploit libraries such as Streamlit or Gradio for quick front-end development, call LLMs through their APIs and language-specific SDKs, download pre-trained models from Huggingface, or exploit tools that facilitate the end-to-end cycle such as Lightning, and achieve something in a short time that is still low-fidelity but includes real data and real AI.  

To add to that phenomenon, Generative AI tools are increasingly being used as prototyping tools in themselves, both by experts and non-experts—the latter benefiting from rising popularity of coding assistants like Cursor or Lovable. All this has contributed to the emergence of the new—and somewhat controversial—term vibe-coding.  

All in all, the skill of AI prototyping demands not only an understanding of how these tools function from a "full-stack" perspective but also the ability to navigate, keep pace with, and creatively manipulate this dynamic ecosystem. Even for technical profiles, it further necessitates a discovery-driven mindset and an intuitive sense for the business needs.  

The impact of easier prototyping

As different companies have different approaches to prototype products, the adoption of prototyping skills might be diverse and progressive. For instance, some data scientists or business analysts, already involved in tasks of contributing to AI product creation, could see how their work shift towards producing and validating prototypes in detriment of model training or creation. Developers or designers with a technical inclination but limited experience in AI may observe the barrier to working with AI components becoming less daunting. 

The increased value of prototyping paints a promising future in which decisions are taken ‘by doing’

What is clear is that making prototyping cheaper gives more value to asking the right questions and to formulating the relevant business hypotheses. Likewise, it blurs the boundary between technical work and the work of taking the right business decisions.  

This blurring could go as far to influence non-technical, managerial roles. A good example put forth by Andrew Ng is that of product managers. As product managers are responsible for prioritizing the product features, assure the customer-product fit and explore with risk and creativity, cheaper AI prototyping increases the importance of their roles. Moreover, there is a growing possibility that product managers might assume some prototyping tasks. This claim aligns with certain perspectives around the need for mixed technical/management roles that were postulated before the rise of generative AI, such as that of the Analytics Translator. I have personally encountered examples in both sides: product managers who have undergone AI training in order to communicate more effectively with their teams, as well as technical profiles who promoted to management roles and still reserve some time to test technical ideas before delegating them to their teams. 

Imagine a future where investing time in a quick prototype could prove to be more effective than a series of endless PowerPoint meetings based on speculations. Or where a design thinking workshop becomes a “design doing” line. Ultimately, the increased value of prototyping enables us to speculate about a promising future in which decisions are taken “by doing” at multiple levels of the organization, displacing other less effective practices.