The community of innovators widely accepted three principles of creating a successful product: user orientation, prototype testing and agile development. The methodologies of design thinking and lean startup brought these approaches up to the level of the “industry standard”. Today, every product team is involved in close interaction with users during the development process. However, our experts have come to understand the fundamental limitations of these principles accepted among community member. In this article we present three key issues of popular innovative methods in the context of artificial intelligence.
The innovation standards
Over the past two decades, the innovator community has defined a set of methods and approaches needed to create new products and services. Designers proposed a design thinking methodology [1, 2], which combined user research, brainstorming and prototypes for testing the ideas into a single process. User orientation has become the key component of the innovation process, allowing the product team to better understand and coordinate the needs of users in the market. Unlike traditional scientific or engineering methods, in the design thinking process a new product is formed gradually, step-by-step. Thus, the team is no longer under the pressure to find the perfect solution here and now.
Not only do the principles of design thinking help in shaping the product itself, but also to define the process of creating services and business models. By combining these ideas the Lean Startup approach  provides entrepreneurs with guidance on how to innovate. In general, these concepts have been so successful that we could describe them as an “industry standard”. Our experts – researchers, teachers and consultants – are also involved in spreading design thinking and lean startup methods among students and industry experts.
Where the “industry standard” does not fit it?
Over the past few years the SAMSONOWA & Partners team has been helping OWN.space to create new services based on artificial intelligence, the so-called digital assistants. In the course of our work, we understood some of the fundamental limitations of the previously mentioned innovative standards. Below we would like to describe three challenges that we faced and thereby start the discussion about the ambiguity of standard approaches to innovation.
1. User research is less relevant
The first step in the process of creating innovation is to determine the needs of the user, in other words, whether the client needs a specific solution. The main task of the innovation process is to solve “user mystery.” In order to do this, innovators use empathic design methods, they immerse themselves in the lives of customers, observe users or cooperate with them to come up with solutions together.
We found that in the context of AI, user research plays a much smaller role. Indeed, we noticed that the search for options of using AI as a digital assistant is not difficult. All options can be basically divided into two groups: human labor automation and big data processing. In both cases AI solutions are in demand and specific application ideas are almost unlimited.
Instead of spending a lot of time trying to understand what users want, it turned out to be much more important to understand technical capabilities and limitations. Of course, the ways in which people solve a specific problem are important to observe when creating a technology (if it can be transferred into an algorithm). However, in many cases, solving the problem with artificial intelligence is very different from the way humans approach work. Consequently, user research is losing its relevance.
2. Prototypes are not very useful
After defining a product concept, design thinking suggests creating a quick and cheap prototype for testing ideas with users. The prototype visually shows the user an idea, an initial concept. The way the client interacts with the prototype allows designers to test hypotheses and make the necessary changes. In addition, the team often understands the technical limitations of bringing the idea to life during the prototyping stage.
In the case of artificial intelligence, prototypes are not so useful due to several reasons. When we work with machine learning models, it makes no sense to use cheap prototypes like “The Wizard of Oz” (when a user’s interaction with the system is tested while another person performs the system’s functions). Although it allows us to easily check the general concept, this prototype tells us too little about the technical capabilities of the real solution.
Technical prototypes, on the other hand, are often very disappointing at first. Indeed, the first outcomes of artificial intelligence technology almost always fail to produce the desired results. For example, Google Translate has been widely known for its mistakes in the past, but everyone recognizes how good it is today thanks to a long learning period.
To get good results in a short time you can focus on one case or scenario. So, for all digital assistants, we could have created an optimal prototype – a bot that would deliver ideal results for a predetermined scenario. However, such a prototype would say nothing about the technical feasibility of the project since one case could radically differ from another.
3. Step-by-step changes do not always lead to success
Iterations and cycles are the basis of design thinking and lean startup. You should state a hypothesis, test it on a cheap prototype and either integrate it into the market or reject the solution. This approach allows you to approach the perfect solution step-by-step and incrementally improve the product.
In our projects, we saw that improvements in the quality of results were non-linear regardless of our efforts. Small changes in technology could lead to sharp increases in the quality of results, or they could stagnate or even worsen the results. Therefore, we could not evaluate the development process as a function of time.
For example, our science agent performed very poorly in the beginning. Changing the approach to training our deep learning model led to a jump in performance. In another case, despite the initial progress, we ultimately reached a plateau and were never able to achieve the desired performance for the agent that monitors business activity in the market. Projects that looked bad initially could still turn around and others that looked promising turned into dead ends.
We wonder why standard innovative methods do not work in the context of artificial intelligence? In the past the scientific community has turned remote research sites into user-friendly, innovative laboratories. We went through a stage when scientists and engineers became empathic design thinkers. Artificial intelligence technologies may require different, more isolated types of innovation. We should be very attentive to such signals and develop our management style accordingly.
- Rowe, Peter G. Design thinking. MIT press, 1987.
- Beckman, S.L. and Barry, M., 2007. Innovation as a learning process: Embedding design thinking. California management review, 50(1), pp.25-56.
- Ries, E., 2011. The lean startup: How today’s entrepreneurs use continuous innovation to create radically successful businesses. Currency.