The AI-driven company: the R&D process

As part of this series on the AI-driven company, we discuss three maturity models that are the primary findings of the interview study we conducted during the spring. We completed the first one, AI in the business processes of the company. Now, it’s time to dive into the second maturity model, which focuses on the R&D process.

We identify five stages of maturity in the R&D process, ranging from the AI assistant, AI compensator and AI supercharger to the AI system generator and the kaizen AI generator. The AI assistant stage is very similar to the first stage in the business process maturity model in that it’s about generative AI acting as an assistant to individual developers. In the case of the R&D process, this means that developers can use AI to autocomplete code that they’re writing, generate test cases and documentation, as well as other smaller tasks that simply make them more productive and allow them to spend their time on the most value-adding activities.

The AI compensator stage is where an AI agent becomes a virtual team member to a team of software engineers. The main value the agent provides in this case is to bring competence to the team that the other members lack. In our interviews, we’ve seen a number of examples. In some cases, it’s concerned with using new tools or frameworks that the team hasn’t used before and where the agent can provide knowledge and support in using them effectively. In other cases, it’s concerned with domain knowledge. The team may be asked to work on an application in a domain where they have limited experience and understanding and the agent can provide details on the peculiarities of the domain. Finally, the agent may support the team with a part of the development process that they don’t have much experience with. For example, when building software subject to regulatory compliance, evidence needs to be collected to show that the software satisfies the regulations and an agent can then provide input on the necessary steps, activities and artefacts.

The AI supercharger stage is where most of the team is replaced with AI agents. In one of the cases that was shared during the interviews, an enterprise architect was teamed up with a set of AI agents that dealt with generating requirements, designing the architecture, generating the code and test cases, as well as conducting integration and testing. The role of the architect was predominantly to provide the right prompts to the agents and to supervise their output. It’s in this stage that the real efficiency benefits of adopting AI become prevalent.

At some point, we even look to get rid of the last human on the team. Then we enter the AI system generator stage. Here, a set of agents receives a prompt expressing the intent of the individual, team or organization with the system that should be created. Based on that intent, a team of agents goes through the entire process of generating the system and provides a tested, documented and complete system as an output.

The final stage is where the AI agent team doesn’t just generate and deliver the system, but rather stays involved and continuously monitors it, experiments with alternative ways of realizing certain functionality with the intent of improving performance and regenerates code whenever necessary to keep the system improving continuously. Of course, for systems where there are multiple or even many instances, the kaizen AI generator combines the data coming back from all of them to provide solutions that optimize across a fleet of products and can provide mass customization.

In many ways, it shouldn’t be surprising that software engineering is the discipline most proactive in adopting generative AI across industry and society. In the end, we work with software technologies all the time. However, in my close to thirty years as a (full) professor in the field, this is the first time my own field gets disrupted. All previous transformations affected other disciplines and industries and resulted in more work and opportunities for software engineers.

One of the concerns is that this will remove the need for software engineers, but I’m not concerned about that at all. Jevon’s paradox shows that when something becomes cheaper, its use goes up dramatically as well. As software engineers aided with GenAI tools can generate today’s systems at a fraction of the cost, the demand for software will only go up. And this will also require engineers in the process. However, you need to have the relevant skills to be able to make use of this!

After the business process AI maturity model, we’re diving into the AI-driven R&D process. We identify five stages that companies evolve through in their R&D: the AI assistant, the AI compensator, the AI supercharger, the AI system generator and the kaizen AI generator. In this series, we focus predominantly on software R&D, but of course, mechanics and electronics are affected as much by generative AI as software R&D. It just takes different forms. To end with a quote from Fi-Fi Li: “Artificial intelligence isn’t a substitute for human intelligence; it’s a tool to amplify human creativity and ingenuity.”