A reflection on the AI Strategy Lab, Team 20: Howden AI Risk Profiling
Can AI turn a loss-making insurance policy into a profitable one for a broker? That was the question behind our AI Strategy Lab project. I was convinced the answer would be yes. However, by the end, the numbers taught me something more interesting than a simple yes or no.
The idea we started with
Our team worked on a case around Howden, an insurance broker. Mid-sized Dutch companies often fall into a gap. Their risks are too complex for a standard insurance package, but are also too small to justify the expensive, labor-intensive process of co-insurance, where several insurers each take a share of one risk. Before a broker can approach those insurers, they must collect and structure documents such as, policies, claims histories, risk reports and photos of buildings.
Our solution was an internal AI tool for brokers. It reads those documents, turns them into one standard risk profile and proposes which insurers on Howden’s panel could take which share. The broker always checks and approves before anything is sent.
In our project proposal, I wrote the disruption rationale: the argument for why AI changes the game here. My reasoning was that generative AI can process unstructured information at a much lower marginal cost. That would cut the time and cost of preparing a placement, which is exactly what makes smaller, complex risks not worth the effort for brokers today. It sounded convincing. It also turned out to be the assumption we had to challenge most.
My role: the big picture and the final call
In the final phase, my part of the deck was the critical evaluation and recommendation: the ecosystem and competition, governance and regulation, a future scenario where everyone uses AI agents and the final recommendation.
I also compared our slides with the nine required components of the official assignment and checked whether our story flowed from one slide to the next. It didn’t, at first. Different slides used different example premiums, so you couldn’t compare them. One slide already gave a recommendation before the actual recommendation slide. The ecosystem and competition analysis, a required component, had disappeared from the deck. I shared these points with the team and we fixed them. The communication within the group went very smooth.
The moment the numbers disagreed with us
The most interesting part of the project came when I looked closely at our economic model to write the recommendation.
Placing one insurance line costs a broker about €368 in time, with a commission of around 15%, a policy needs a premium of about €2,450 just to break even. With our AI tool that break-even point drops to about €2,275. That is an improvement, but a small one. A €2,000 policy still loses money, even with AI.
Honestly, this surprised me. I don’t have background in insurance; I expected AI to make small policies clearly profitable. It didn’t. Even for larger policies the gain wasn’t really in money but in time and even that gain was smaller than I had imagined.
So, we had a choice: ignore the numbers and keep our original story or change the story. We changed it.
From “unlock a market” to “free up time”
The real value of the tool turned out to be broker capacity. In our base case, it frees up about 1,610 hours a year, roughly one full-time broker. That matters because the co-insurance market has a staffing problem: experienced people are retiring faster than they are replaced. The scarce resource isn’t the policy it’s the broker’s time.
That shift shaped our final recommendation. We didn’t recommend a full rollout. We recommended a three-month pilot with a clear stop rule: the tool needs to save at least 22 minutes per insurance line to pay back its costs in the first year. If it does, we scale. If it doesn’t, we stop.
Learning from what went wrong
Not everything worked and some of the most useful lessons came from that. While reviewing our prototype, we noticed an error. For one client, the insurer shares added up to €8.4 million, while the tool advised €9.2 millions of cover. Yet the screen still said, “full cover”. We used this in our governance slide as a failure example. It showed exactly why a human broker must check and sign off on every file.
Teamwork and Demo Day
Our collaboration went smoothly. We split the work into clear packages, communicated well and everyone delivered their part and adapted it to the shared layout of the deck. Working in separate parts made one thing clear to me: someone needs to look at the whole, otherwise four good pieces don’t become one good story.
Demo Day was a fun and interesting day. I presented our idea together with my group during the rotation rounds. What I enjoyed most was seeing how other students had worked out their ideas and especially, seeing their working prototypes. The small audiences made it personal.
What I’m taking away
Three lessons stay with me. Firstly, cheaper is not the same as valuable. Before this project, I mainly saw AI as something that makes work faster and cheaper. I now see that a cost reduction doesn’t automatically create a business case. The question is where the value ends up and who captures it. In addition, when everyone has AI, AI is not your advantage. Companies like Cytora already sell document-reading AI to insurers. If everyone can buy the same tool, the advantage moves to what can’t be copied, such as: relationships, data trust and accountability. For Howden, that means its insurer panel and its brokers. Lastly, a good strategy knows when to stop. The strongest part of our recommendation isn’t the idea itself, but the test. One clear number, a short pilot and the willingness to walk away if the evidence disappoints.
I started this project by writing why AI would disrupt a market I finished it by learning to question my own argument. That might be the most useful thing I have learned about AI strategy so far.
