Project Learning Blog – Felix Brugger, Team 12

No ratings yet. For the AI Strategy Lab, our team developed ImpactIQ. The track B project consists of an agentic AI for small and mid-size general contractors in the construction industry. Procurement makes up a major part of a construction project’s spend, and earning margins are slim, so there is little room for surprises. Our premise was that when a supplier announces a delay, pricing its impact means manually linking several documents, which costs hours of manual work. Consequently, small to medium size contractors only price big supply chain disruptions, even when small disruptions hold the potential for causing significant costs. The aim of our agentic AI is to take over this evaluation step. It can read messy supplier data, it prices three possible responses, and recommends one, all while a human makes the final decision.
We decided to split the work between us so everyone could contribute equally to the project. I was responsible for the first two points, which were problem definition and market opportunity. Essentially, my responsibility was to show why the identified problem mattered and why it could be solved at this point of time. Each of us wrote out our section in a shared document, and other teammates then turned the content into slides.
I gathered industry figures and developed our core insight, namely that small supply chain disruptions go unpriced. Instead of describing our agent’s features, I chose to break the contractor’s workflow into three steps, which showed exactly which step our agent would take over. To argue why the timing is right, I explained how generative AI can now turn messy supplier data into valuable insights. I then used the newly vulnerable market framework to show why this market is the right one to enter. It is newly easy to enter because of lower entry barriers, attractive to attack, and hard for incumbents to defend.
As my section came first, much of the work of my teammates that followed built on it. To me this made two arguments feel like a bigger responsibility than what they may look like. I feared that if the problem wasn’t defined precisely enough or the underlying argument wasn’t articulated well, nothing built on top of it would be convincing.
For me, the most important decision was how much the agent should be able to do. I initially liked the idea of an agent that not only evaluates a disruption but also decides on the response by itself. However, the more we discussed it, the less sense it made to me since a wrong call on a construction project is expensive. We therefore limited the agent to recommending only and left the decision with the project manager.
In hindsight, I see this decision more as an important point for our value proposition rather than failure early on. Through the agent the manager is able to make better informed decisions and, importantly, doesn’t get replaced by it. We decided to keep the human in charge where there’s weak evidence coming from the agent.
Our strategic positioning went through a similar pivot. At first we wanted ImpactIQ to disrupt the construction industry, but this was too ambitious for a small venture and would have required far more industry knowledge than we had. Instead, we positioned it as a complementary innovation that augments a human’s work rather than replacing it. I believe this made our story more credible.
The biggest surprise came when we had the prototype up and running. Up front, I assumed that more data, regardless of its kind, would translate into more accurate estimates. A mistake from my side was not to question this assumption, even though my own part relied on it, as I had argued that cross-project data would make us hard to copy. When the prototype was trained with data from similar projects, it cut the estimation error by up to 44%. Conversely, it showed that when the underlying data weren’t that similar, it actually made worse estimates than ignoring the data altogether. My main conclusion was that more data only provided additional value when it was actually similar enough. This taught me that a prototype’s real value can often be in proving us wrong.
The feedback on our project proposal was positive, as our problem was narrow and precise enough from the beginning on. We did not need to pivot, and it confirmed that a specific problem like unpriced small disruptions is easier to build on than a broad idea. Instead, my group and I got a lot of feedback from the working prototype and changes to our project were mostly derived from it. For example, we decided to add a comparability threshold under which the agent would decline to provide a cost estimate when its reasoning is too weak.
Personally I found time to be the most difficult part of collaborating. Some of my teammates work alongside their studies, and others don’t. This sometimes made it difficult for us to find a day when all four of us were available. A learning I made through this is to not wait for everybody to be available, but rather to just move on with the people who can make it that day. However, this only works when the other teammates are kept in the loop. Thus, communication became much more important.
Before this course, I largely saw AI as an inherent advantage but this project showed me otherwise. My “why now” argument was that GenAI can now cheaply read messy supplier data. However, every competitor has access to the same models, so the capability that opened the market also makes our translation step very easy to copy. What remains is something AI cannot generate, namely a record of real disruptions, the responses chosen and what they actually cost. I concluded that using AI alone doesn’t make a business flourish and that how a business applies new technologies is what actually counts. A discussion in class on software development made this clearer to me. When generative AI arrived, many developers feared for their jobs, yet we discussed companies that replaced developers with AI and later had to hire them back, because applying AI in practice proved much harder than expected.
Our project showed a similar logic on a smaller scale. The agent turned out to be most useful when it was carefully designed to recommend rather than decide on its own, and to stay quiet when its data was too thin. For me, the question now isn’t anymore whether a business uses AI or not, but where it adds value that customers will pay for and competitors cannot easily copy.