Introduction
For our AI strategy lab we developed an app that helps early-career researchers find the right expert beyond their own network. The app is called FieldFinch and is an AI-driven personal matching agent based on professional capabilities. With the app we wanted to lower the barrier for researchers to connect with each other and efficiently find expertise on certain subjects outside of their network.
How feedback changed our direction
Our first proposal was far too broad. We wanted to be able to match project members, colleagues, students and employees. During the lecture, we were told that we had identified a real pain point but had not tied it to a specific user, decision or workflow. Looking back, I agree that we were thinking too broadly. This is because we started from what AI could do, match people on capabilities, instead of from whose problem we were solving. Narrowing to early-career researchers made the problem concrete: 76% of researchers find collaborators through their existing network (Borromeo et al., 2014), and early-career researchers with small networks are hit hardest.
My role and contributions
My main contributions were the business model and the governance and failure analysis. For the business model, I worked out how FieldFinch creates, delivers and captures value. In our proposal, we imagined the matching agent as a feature inside LinkedIn and assumed value would be captured through higher user engagement. When we narrowed our focus to early-career researchers, that logic fell apart. The people who benefit most from our app, early-career researchers with small networks, are exactly the people who cannot or will not pay for it.
This forced me to separate value creation from value capture. FieldFinch creates value for four groups. Researchers save search time, experts gain visibility and collaborations, institutions get more interdisciplinary work and stronger grant applications, and companies get faster access to academic expertise. For the pricing method I used the theory of session 5 of the course that described the pricing of platforms and the strategies for two-sided markets (Eisenmann et al., 2006). I decided that researchers and experts use the platform for free as a subsidized side. Universities pay an annual licence, and companies pay per accepted match or a commission on paid sessions as a charged side. Researchers and experts are subsidized because their participation creates the network effects that attract paying universities and companies.
Our governance and failure analysis showed where FieldFinch could go wrong. A language model can write a convincing but false explanation of why an expert fits. A vague question can be misread, and well-known researchers could keep rising in the ranking while early-career researchers stay invisible. Instead of seeing these as weaknesses, I used them to shape the design. The AI turns vague questions into concrete requirements, searches across institutions and fields, and explains each match with links to real publications. Ultimately, people make every decision that matters. The seeker reviews the requirements before the search runs, experts keep control over their own profiles and decide whether to accept a request, and only completed meetings can be rated. Researchers can opt out, and requests are processed within the EU and not used to train the model. Thinking through these risks became one of the most valuable parts of the project for me. Rather than trying to make the AI perfect I combined what AI does best with what people do best, and the guiding principle became simple: AI recommends, people decide.
What surprised me
What surprised me most was our evaluation. We tested FieldFinch on five real interdisciplinary collaborations, and it found five times more experts from other fields than a keyword search, at similar relevance. Before testing, I expected the AI to win mainly on accuracy. A well-written keyword search, however, already reached 68% relevance, so the real strength of our approach turned out to be reach, finding relevant experts in fields a researcher would never think to search. Our second version taught us just as much. It raised cross-field reach from 20% to 44%, but none of its suggested experts were on-topic anymore, so we decided to keep the first version. Reach without relevance helps no one. This showed me how important it is to test every change against a fair baseline, instead of assuming that a more advanced setup automatically works better.
Collaboration
Our collaboration went smoothly, largely because we invested time in brainstorming together. This made sure that we were all on the same page to write the proposal. Each of us owned a part, which gave everyone clear responsibility. What I learned is how much these parts depend on each other. My business model only made sense once the target group was narrowed down. The pricing depended on the cost estimates, and the ecosystem analysis influenced the failure analysis and the final recommendation. The most valuable moments were when we collaborated and worked together to connect those parts and created a sound strategy.
How the project changed my understanding of AI and business strategy
The AI strategy lab changed how I look at AI and business strategy. At the start of the project, I thought a good model would make a good product. However, I learned that the value of an AI product does not come from the technology alone, but from how it is tested, designed and embedded in an ecosystem. Our evaluation showed that AI’s strength lies in one specific area: reaching across fields. The failure analysis showed that a trustworthy product lets AI recommend and people decide. Because language models and publication data are available to everyone, I now believe the lasting advantage of an AI venture lies in what grows around the technology. For me, the biggest lesson is that a good AI strategy does not start with the question of what the technology can do, but with the questions of who it helps, who pays for it, and why people would trust it.
References
Borromeo, C. D., Schleyer, T. K., Becich, M. J., & Hochheiser, H. (2014). Finding Collaborators: toward interactive discovery tools for research network systems. Journal of Medical Internet Research, 16(11), e244. https://doi.org/10.2196/jmir.3444
Eisenmann, T., Parker, G. & Van Alstyne, M. (2006). “Strategies for two-sided markets.” Harvard Business Review, 84(10).
