My role in the project was shaped by the fact that the idea the team chose was mine, for Howden Netherlands, an international insurance broker I have worked at since my bachelor’s. Because of that, I became responsible for the internal information: how long the current workflow takes, who is involved, what it costs. I had to ask operational colleagues on the floor for input and whether the idea would work. The first reactions were positive, which was useful validation but not evidence on its own.
I was also responsible for building the prototype. I approached this from a functional angle: what does the broker need at each step, where is the bottleneck, how does the search for coverage and pricing actually work. My experience in the sector helped here. I had already known for years which sources the sector sees as authoritative, for example KPMG reports and VNAB standards. This also helped me in signaling inconsistencies in the deck. A source was cited to claim that Howden had a large number of M&A policies, as support for a co-insurance argument. M&A policies are however not a form of co-insurance policies. This is not something you would easily know unless you work in the sector, which gave me a chance to redirect our course.
This was also the first time I worked in an academic setting since taking a gap year between my bachelor’s and master’s, which meant I had to learn the dynamic in such a project again. As a team we did not have much disagreement. I think part of the reason was that I had experience in this particular sector we chose, so decisions on market facts went through me by default. That is useful for accuracy, but it also means the team agreed quickly in places where someone with more insurance background might have pushed back. If I did this again, I would deliberately ask someone to challenge my input before we built on them, instead of letting agreement follow automatically based on who has more experience.
The prototype decision
Testing the prototype properly was never straightforward, mostly because we were not allowed to use real client data. I built a set of fictional documents instead, matching the format of real policy files as closely as I could, which was the right call for confidentiality, but it also meant that any “evidence” from testing came from documents that would be more standardized then a real file.
That same limitation runs through the other big decision in the prototype: whether it would be a working AI agent or a prescripted flow. Nothing was fixed about this at the start. We eventually chose prescripted, and that looked final until later on, when I actually built and tested a live version. But because that version was also only tested against documents I had constructed, it never really settled how it would hold up on a real file, which is part of why we eventually chose the prescripted version. Building it meant using Claude throughout, for a project about how firms should use AI responsibly, and the gap in how much correction it needed was telling: generic, interface-level work needed almost none, while anything specific to Howden, to the tools’ technical limits, or to the fact that our test data was something I had built myself needed real correction. We say as much directly in the deck: the business case still rests on assumptions, not on anything measured against real files.
That gap sat next to a second one: we never had an outside check on any of it either. Early in the course we had the option to submit a short project proposal, which was meant to flag problems before we went too far. We did not get feedback on ours, which would mean we were heading the right way. That is a reasonable system, but it meant the project ran on the assumption that silence equalled approval. There was also no external review of the prototype itself: no underwriter, client or outsider looked at it before Demo Day. Put together, the positive reactions from colleagues early on, the team’s quick agreement on market claims, and this silence-as-approval all did the same thing: they felt like confirmation without ever being evidence that would hold up outside the room where we produced them.
The one point where something closer to real testing happened was when I stress-tested the prototype myself and found an inconsistency in the numbers on one screen. It was not proof a live AI system would make that mistake, since the content was prescripted rather than generated from real analysis, but it showed something polished can still be wrong without anyone noticing.
What changed in how I think about AI and strategy
That absence of outside scrutiny is also what reshaped how I think about AI and strategy. Before this project, I thought of AI mainly as a threat to brokers because it removes the information gap that justifies their role. A broker like Howden exists because it reduces the asymmetric information between a client and the insurers assessing its risk, the electronic brokerage effect Malone et al. (1987) describe when a market no longer needs an intermediary to bridge that gap. Working on the deck pushed this further: if clients and insurers also run their own AI agents, the time-saving part of our tool, reading and structuring documents, becomes available to everyone, not just Howden, so the asymmetry it closes gets closed just as easily by someone else’s tool. That is disintermediation in its sharper form, where the value does not shrink but moves to whichever side self-serves first.
The answer to that is not a faster pipeline but becoming the party both sides route through regardless, the shift from pipeline to platform that Van Alstyne et al. (2016) describe: competing less on processing speed and more on being the standard both insurers and clients check against. Co-insurance already has this structure, with Howden between companies and an insurer panel, so AI mainly raises the question of who runs that intersection once reading documents stops being where the value sits. What is harder to copy in that version is the verified file a broker is accountable for, and the panel relationships behind it, the kind of complementary asset Barney and Reeves (2024) argue remains defensible even when the AI capability itself is not. AI does not remove the information asymmetry that justified Howden’s role, it relocates it, from who can read a file fastest to who insurers and clients trust without checking it themselves. That trust is exactly what our own prototype never earned: a tool is only as valuable as the confidence others place in it without re-checking, and ours was never tested by anyone positioned to grant or withhold that confidence.
References
Barney, J. B., & Reeves, M. (2024). AI won’t give you a new sustainable advantage. Harvard Business Review, 102(5), 72–79.
Malone, T. W., Yates, J., & Benjamin, R. I. (1987). Electronic markets and electronic hierarchies. Communications of the ACM, 30(6), 484–497. https://doi.org/10.1145/214762.214766
Van Alstyne, M. W., Parker, G. G., & Choudary, S. P. (2016). Pipelines, platforms, and the new rules of strategy. Harvard Business Review, 94(4), 54–62.
Student: Raoul Madijokromo
Studentnumber: 664528
