As part of the Information Strategy course, I formed a group with three other students to drive back the high online returns processing costs for Adidas while transforming its webshop’s customer experience. Our solution involves AI-driven sizing recommendations embedded into the webshop when shopping for clothes. Besides my (content) contributions in the idea generation, evaluation and future market tests, I also took on a managerial role to guide my group the right way in the process of idea generation and problem identification. The AI strategy lab journey taught me that true business value is not achieved by an AI model itself, but more so through its proprietary feedback loops and customer trust.
My core contributions within the group’s work revolved around problem specification, governance & failure analysis, future market test and the implementation of our tool. At the time of our proposal, we had not specified the business problem we were trying to solve enough. Although the problem (high returns processing costs) seemed evident, we did not have the hard figures and arguments to back up the need for a solution/proposal. Although the problem specification was a group effort, I went to extra lengths to pinpoint the exact economic cost of the problem to Adidas (€690 M annually) using reputable scholarly sources and Adidas’s financial statements.
I also took lead on the future market test of our product. To be more precise, how does our product stand in the market when direct competitors such as Nike & Puma and third-party retailers like Zalando get access to similar AI models? One of the IS course components I learned is that value is not created or captured by an AI model itself – the competitive advantage comes from proprietary data integrated into feedback loops. So, I suggested we design our AI sizing tool to create a proprietary data flywheel using Adidas’s dataset of historical purchases, returns and product reviews. This way, any new sale, online return or product review improves our tool’s sizing recommendations.
Lastly, I took ownership of the governance & failure analysis part of our project in two ways. Firstly, I addressed the data privacy issue of our model. Under EU law GDPR article 5.1, body data meets a specific sensitivity threshold which means the data cannot be stored without consent. I suggested the opt-in consent feature to meet the European data privacy law. Secondly, I created a fallback mechanism for our sizing tool. Should the amount of returns with cause size/misfit exceed 5% of the baseline for that product, our system will fall back to traditional sizing charts until the anomaly has been evaluated and solved. This prevents AI-hallucinations and systematic mis-sizing.
The first important decision we had to take was which specific business problem are we going to tackle? Our initial proposal, an off-the-shelf AI-powered dashboard that analyzes return patterns and generates actionable recommendations for SME online apparel retailers was not specific enough. This idea revolved around multiple problems: high return processing cost & lagging AI-adoption within SMEs. We assumed that hitting two distinct business problems would help us disrupt the market, but we realized not much later that our idea was too broad and the available data was fragmented and lacked depth. So, we took a step back to the drawing board and decided to specify just one solution: the high return processing costs to a specific online apparel brand: Adidas. This pivot didn’t only make the project more tangible, the problem now also was specific, and more importantly, measurable.
A second important assumption we had to address was the dependence of our AI-tool on truthful consumer input; after all, the model will make recommendations based on the physical traits (such as height & weight) a customer fills out on our website. Should consumers (out of shame for instance) not fill out this information truthfully, the size recommendations of our tool would be off. I suggested that we address this issue by working around this potential consumer behaviour instead of trying to fight it: we decided to present customers with a percentage match per available product size on the webshop. Should the highest match still be lower than 60%, our model would ask a follow-up question to improve accuracy of its recommendations.
Despite not being direct feedback to our project, the first lecture after the proposal deadline made it clear that our proposal was still too broad and did not focus on ONE specific business problem. Professor Li challenged our idea (though not directly) when she mentioned the strategy lab should revolve around one specific business problem, preferably one we encounter ourselves. This was surprising, as we initially had a pretty high confidence in our business problem. The challenge of finding data and evidence to back up what we thought to be a valid business problem and its (economic) need, showed us our business problem was not yet specific enough. Therefore, we decided to get back together and specify the ONE business problem we were trying to solve first, backed up by valid numbers as evidence. The evidence we found supported our assumption that the high returns processing costs are a business problem in need of solving. This led us to our pivot to develop an AI-driven embedded sizing tool within Adidas’s webshop.
My three fellow team members were already working together for the other course as part of BIM. Being the ‘outsider’ in the group allowed me to quickly identify and analyze my teammates’ pre-developed way of working. It became clear to me that all team members had strong ideas – but the idea generation discussions were lacking direction as everyone was focussed mostly on pushing their own ideas, but not really listening to other ideas. Therefore, I tried to take on the role of discussion-moderator to shape our discussions so that everyone was able to explain their ideas without being interrupted. This worked quite effectively – we went from discussing all three ideas simultaneously to discussing one problem at a time, at a much deeper level. It taught me that it can be valuable to sometimes observe my surroundings a bit before being able to engage in a meaningful discussion: there were enough strong ideas, we were simply not evaluating them properly.
My view on AI’s role in business has changed significantly throughout the AI strategy lab project. I was aware of many of the different possibilities that GenAI and Agentic AI have created. However, I am now much more informed of how the real value to businesses doesn’t necessarily lie in the AI model itself, but in its integration and usage: creating a data flywheel can lead to every new datapoint improving the overall quality of an AI-model. The real value is captured when data governance, consumer trust and human-in-the-loop safeguards are met as conditions.
