Project Learning Blog

Introduction
During our project, my team and I designed DIPS, or the Demand & Inventory Planning System. We created DIPS to add external trend signals to LOAVIES’ inventory planning so planners can make better reorder decisions. As my role, I built and hosted the prototype, designed and ran its evaluation, and wrote the future-market test.

How we arrived at LOAVIES
When we started with the project, we had one goal in mind. We wanted to solve a real problem. We initially wanted to design a trend-based demand forecasting platform sold as SaaS to fashion retailers, which meant working on a new venture. Thankfully, through feedback in the early stages of the assignment, we learned that a generalised forecasting tool would not let us test our core assumption, that “external data adds predictive value beyond what historical sales already provide”. As a result, we decided to switch to implementing changes into a real organisation. Our first targets were large retailers like H&M and ZARA, but they already had advanced planning systems, so DIPS would add little value. We therefore turned to SMEs that lack the money or technical expertise for existing AI forecasting tools. After a long search, I proposed LOAVIES, because of their openness to using AI, and weekly drops of new products that make them highly exposed to fast-changing trends.
Our first real setback arrived when LOAVIES could not cooperate with us because they were busy with the fall drop. Without company data, our business case was weak. Still, the fit with our idea was too good to give up, so I looked for large public datasets with similar products and information. This decision was probably what shaped everything I did afterwards.

Our decisions and iterations
At first, we imagined an agentic system that could replace human oversight. Reading about AI in organisations (Wilson & Daugherty, 2024) changed our stance, however. Companies need AI that augments people, not AI that removes them. We also received feedback to keep the technical design simple to fit SMEs as our target, who need something cheap and easy to adopt. So, we built DIPS as decision support instead. It advises, and the planner makes the final decision.
The first version of the prototype let the LLM calculate and classify the trend and capped the trend effect at ±15%, and left “confidence” as a value undefined. When I tested it against the decision rule, it scored STOP, which means it prevented at most 18% of lost sales, just below the 19.3% break-even we estimated in the economic analysis. This unfortunately meant I had to revise the prototype. All calculations and actions got moved into code, so every run gives a consistent answer, and the LLM now only explained the advice. This was an important lesson that an LLM is not automatically the best or right tool for every step of an AI system.
Because of how our model works, there were no real sales after the decision week, which we had to model. The key unknown was “s”, the share of a trend change that turns into real purchases. Due to the nature of the data, instead of choosing one value, I tested the full range from 0 to 1 to ensure consistency.

What was surprising, and what failed
The biggest surprise was that the second version was also not “perfect”. It passed five of six checks but scored REVISE on the decision rule, which meant that when a trend was pure hype (s=0), leftover stock rose by 36%. What is even more interesting is that forecast error fell from 7.8 to 4.5 units, while leftover stock was still high. Before this project I assumed a more accurate forecast was automatically a better one. Now I understand that accuracy is not the same as value. I tested two quick fixes, and neither worked. One barely blocked anything, and the other cut leftover stock but pushed prevented lost sales below break-even.
The future-market test taught me a similar lesson. My first fix for fake hype made DIPS wait two weeks, which undermined its main purpose of helping with short-term decisions. I reworked it into a “confirm-in-days” design. DIPS was now proposed to weigh sources by how hard they are to fake, confirm trends with LOAVIES’ own add-to-carts and orders in steps. The last failure came just before the deadline. I noticed the website still showed an earlier version. I fixed it, but left the name accidentally unchanged, a strong reminder to check deliverables earlier.

How feedback and evidence changed our work
The second round of feedback from our professor pushed us toward simplicity and honesty. In the economic analysis, we replaced our own estimates with values from research, which made the business case more solid and defensible. We also continuously reviewed slides from each other to point out where our slides were unclear. In my case, we found that my fake-hype example explained nothing to a listener, so I rebuilt it around one concrete number.

Collaboration
Splitting tasks early on made us work efficiently, but at one point, we became too independent. When we assembled the slides, our sections did not fully match. Changes in my evaluation, for example, affected the cost and break-even figures in the economics section. We solved this with rounds of mutual review where we told each other what to needed to be changed or removed. This was a great reminder, so that in future projects, I will work in short integration-review cycles rather than only iterate at the end.

How my understanding of AI and strategy changed
This project definitely changed how I see AI in business. I started out thinking the value is in being the most advanced, autonomous system. I now see AI mainly as a way to improve human decisions, where judgment and accountability stay with people. I also learned that access to AI rarely gives a lasting advantage, since competitors can buy the same tools. The real advantage comes from internal data and feedback loops that learn from real outcomes. My key takeaway from this project is that when implementing artificial intelligence in complex workflows like inventory planning, sometimes less is more, and we need to stick to our scope if we hope to provide value through our project.