Role & contribution
Throughout the process of shaping DIPS, my role was to uncover the problem and opportunity and design the business model. I have been interested in sustainability and fashion for a long time, and the EU’s new restrictions on destroying unsold clothes caught my attention. This is a problem for fashion retailers as it places them in a difficult trade-off: too little stock means missed sales, while too much leaves them with unsold products and difficult end-of-life decisions. Things get even messier when an asymmetrical burgundy leather jacket trends one day while the next day everyone is going crazy over the newest viral 2000s-style military jacket. So, upon further research, I discovered the need for something like DIPS: an AI inventory management tool that combines historical sales with external fashion trends.
The way I envisioned DIPS to create value is via earlier demand trend signals that ultimately turn into planner recommendations concerning replenishing. Because DIPS would improve LOAVIES’ existing inventory workflow, I framed it within the Existing+ business model, where AI supports an existing workflow rather than replacing it.
Looking at my personal contribution to the team, I believe I unintentionally fell into a leadership role. If work needed to be done, I’d push for it. If a teammate was struggling with their part, I’d take on a supportive role and help them out. We also regularly questioned DIPS’ realistic feasibility. Nevertheless, I trusted our vision and reassured my teammates that, as long as we understand what exactly we are trying to achieve and link it to the course concepts, we were on the right track. However, towards the deadline, I discovered some inconsistencies between our own individual parts. I became anxious about our alignment and pushed for us to review the sections together before submission.
Key decisions, feedback & iterations
Our biggest turning point was the feedback to our project proposal, which served as a much-needed reality check. We initially saw DIPS as a multi-agent system that could access a company’s CRM system and auto-restock based on trend suggestions. In response, this idea was criticised for being too broad and lacking an actual testable assumption. Following up on the feedback, we narrowed our scope towards a single retailer and a core assumption: Can incorporating external fashion trends into historical sales data improve LOAVIES’ demand forecasts and inventory decisions?
To test it, we developed a trend-on vs trend-off comparison with a baseline of a four-week average of historical sales, where DIPS adjusted using Google Trends signals. Evaluation relied on forecast accuracy (MAE), followed by inventory outcomes and economic value – and the evidence gave us a glimpse of hope. Turns out that not only did the forecast accuracy improve – MAE decreased from 7.8 to 4.5 units in our simulated scenario – but we also noted a positive modelled business impact via DIPS – 25.3% of potential lost sales were prevented. Nevertheless, we also noticed an unexpected trade-off from a 17.9% increase in leftover inventory, as well as a failure scenario due to trends that are pure hype rather than actual demand, increasing leftover inventory by 36%. Additionally, only five out of six success criteria were met, leading to a ‘revise’ recommendation rather than immediate implementation.
The test results further nudged us towards rethinking AI autonomy. Our original approach involved agentic AI that automatically executes inventory replenishment decisions. However, our findings support human review, especially under a pure-hype trend. Therefore, our revised approach repositions DIPS as a recommendation generator rather than an absolute source of truth, and places human judgement at the heart of the replenishment decisions.
What surprised me or failed?
When we found LOAVIES, it fit our concept like a glove: a Dutch-based online fashion retailer that introduces new styles each week. Unfortunately, we had no access to their internal data – including sales, inventory capacity, financial statements and markdowns. We reached out to request a call, but LOAVIES declined to participate in our academic project. In the meantime, we were also having a rough time at first making sense of the academic feedback, redefining our scope and business model. As we were stuck in this grey area of confusion, we pushed a lot of conversations about the future of DIPS further down the line.
The resolutions began with Bruno. He showed up not only with a fully functional prototype, simulating a sales database built around 30 LOAVIES tops, but also with a great understanding of how the technical side blends with the business aspect. I believe that was the catalyst we needed to restart our momentum. My personal takeaway here is: next time, when real data is out of reach, I should advocate for building something testable early – like a dataset or prototype – instead of waiting for access to magically appear.
What I learned from collaboration
I got to work not only with good friends, but also with great teammates. Bruno brought strong technical expertise, Federico did a killer job for the financial part, and Carla showcased a critical, structured approach to evaluating DIPS and its implications. Naturally, we worked in different ways, at different paces, but what we had in common was the willingness to help each other out and overcome all the moments when we felt stuck.
Our collaboration shone most during the final week before the deadline. We met every day and went through each other’s slides together to tackle any inconsistencies and make sure our contributions told one coherent story. For example, our replicable assets section still referred to agentic architecture, whereas our final version of DIPS relied on human-approved recommendations. We corrected the wording to reflect our final approach. This made me realise that focusing too much on getting my part done right may lead to overlooking other details that need checking. Next time, I will initiate reviewing our work much earlier.
How the project changed the understanding of AI business strategy
At the beginning of this project, I associated AI’s business value with better predictions and greater automation, so I assumed that improving demand forecasts would naturally lead to better inventory decisions. Our modelled scenario then challenged my assumption. Our findings were that DIPS essentially reduced lost sales at the expense of accumulating more unsold stock. Next, the pure-hype scenario questioned the success of automation even further, as misleading trends could trigger unnecessary purchases. What surprised me was that even human planners missed one out of three incorrect recommendations during our small-scale test. So, while keeping humans involved seems like the sensible approach, it clearly doesn’t eliminate the risk of making poor decisions.
Shaping DIPS made me realise that AI strategy is not purely about finding ways to automate everything, but about understanding where AI can genuinely improve a business process, whether the benefits outweigh the risks and who should ultimately remain responsible for the decisions being made.
