In this project I contributed mostly to the problem statement, the cost and pricing structure and economics. I played the devil’s advocate in meetings, trying to come up with critical questions that people might have when they seeing the project for the first time. I tried to be as critical as possible without being annoying.
Because of my role, the team was able to critically evaluate key parts of the project, like the differentiation strategy. I would come up with a question to one of the teammates parts and I would keep asking questions so they would think of their parts very critically. This led to two benefits: I gained a deeper understanding of their topic, and they rethought their work, often improving the project. This created a team spirit of critical thinking, while we also had fun along the way.
Our first idea was to create an AI dashboard with actionable insights on return data for small clothing retailers (SMEs). After the proposal we sat together and found evidence that such tools already exist. Even though we thought SMEs were a very interesting industry segment, we decided to shift focus from SMEs to bigger retailers, because these companies are incumbents in their industry, with access to millions or billions of data points. With these firms there is often a need for digital transformation and disruptive innovation.
Our second shift was when we decided that prevention was better than optimizing return handling. We found out that return handling is an industry that is highly optimized. We tried to approach fashion returns from a different angle. Retail return prevention is not as advanced yet. We saw this as an opportunity for fashion retailers, where returns are historically the highest. So, we aimed to change this. We focused on optimizing the existing return prevention workflow for Adidas.
The third key decision we made was how to structure our pricing. Our decision to focus on large retailers instead of SMEs meant that we had to rethink our value capture logic. Rather than charging upfront transaction fees, we saw the need to subsidize the customer to build critical mass. This would lead to same-side network effects which we could use to secure the proprietary data required to train our predictive models. This is because customers would choose to adopt AI in the already existing workflow as it would make their shopping more efficient. Subsidizing the customer does mean that someone else pays for it. Adidas pays for their own infrastructure, meaning their investment in the tool is the ‘payment’. In our case the revenue for Adidas came from cost savings, and a more indirect value for Adidas comes from new customer data. The value is captured through cost savings and revenues from targeted marketing.
Initially we treated the AI model as our core differentiator, because few companies seemed to focus on prevention and we believed that we could capitalize on this first-mover advantage. Our competitive analysis showed otherwise: Nike, Zalando and TrueFit already offered similar sizing tools. In the next meeting, playing my role as devil’s advocate, I pointed out that we had no real differentiation strategy. The team disagreed, but our discussion helped shape our idea of the data flywheel as a business strategy: Adidas’ proprietary data, combined with a feedback loop compounding over time, is what a competitor cannot easily copy. That is because for this model we propose that Adidas restructures all their data to one central hub. This is a costly investment, but it is something that competitors cannot easily copy, due to the large investment, especially if they don’t have proprietary data. We ended up changing our value creation from first mover advantage of an AI model to an advantage through data and positive feedback loops. The data flywheel improved our strategic argument, however it remained a hypothesis requiring testing. Next time I won’t assume that a gap in our own research is a market gap. From now on I will look for contradicting evidence, before I deep dive into a project.
My role as devil’s advocate worked well for the team, but it also revealed something about me: when the project got difficult, I wanted to control the outcome. When I said in a meeting that we had no differentiation, my teammates were not convinced, and I felt frustrated because I thought the evidence was clear. I noticed that I was inclined to keep pushing until people agreed. Instead, I mentioned where my disagreement was and together we found alternatives. This led to a more complete argumentation in the project, an outcome I would not have come up with alone. From now on I try to separate challenging an idea from needing to win an argument. I make my point once, with impact, and then, with the team, start a discussion.
We tried to test our assumptions with evidence. We looked for studies that had shown that the logic of our business model worked, but not for studies that showed it didn’t. This meant that, as a group, we failed by looking at our business model as an unchangeable fact because of confirmation bias. We treated competitor gaps as evidence of an opportunity without considering alternative explanations. If we had done this, it would have revealed the model’s limitations and indicated whether we should revise or keep the product as it was. Our business case became a hypothesis presented as a conclusion. However, this failure was a valuable learning experience. In the future, I will approach a project like this one as follows: I will 1) research the market and identify competitors. 2) Define assumptions that are the most critical to our business model. 3) Test technical feasibility using synthetic data when appropriate. 4) Test usability and customer demand with relevant users in a pilot. 5) Evaluate the business case against measurable criteria before committing significant resources. I will do this because in high-risk environments, it is crucial to minimize cognitive biases before investing, as failure can lead to monetary losses.
Before this project, I thought a business strategy for AI meant creating a smarter model than competitors. This project changed that in two ways. First, when we found that our competitors already offered similar tools, our supposedly unique model turned out to be easy to imitate. I now see AI models more as commodities than differentiators. Differentiation comes from adjusting your business model to its capabilities. Second, I learned that a convincing business model is not the same as an effective business model. Never piloting our assumptions made me realize our argumentation was missing a crucial element, research and piloting together create a stronger form of evidence. From now on I will look at a business strategy as a set of hypotheses that need to be tested. I will ask where the differentiation comes from, how the value is created, delivered and captured, and I will ask for both simulated and research proof that it works. Then, I will start a discussion with the team to find a solution rather than trying to win an argument.
