What we did and my role
Street Turn is our venture idea for the Port of Rotterdam and its supply chain that moves through the port of Rotterdam. When an import container is emptied, it usually goes back to a depot. This is while a nearby exporter drives to that same depot to pick up an empty box. Two empty truck trips where one transfer would do. Street Turn reads the messy emails, PDFs and EDI messages where this information lives (anyone in logistics will tell you how messy). To solve this, it hands these messages and emails to a rule-based solver that checks whether a direct transfer is allowed. If it is, a dispatcher gets the proposal and decides.
I owned the competitive and ecosystem analysis plus the governance and failure analysis. That meant working out who controls what: the shipping line owns the box to the forwarder that holds the booking. It also meant asking what goes wrong once a dispatcher starts trusting our suggestions and be very critical on this.
What changed my mind on AI during the course
My first concept table finished with an honest but unhelpful line: “we must decide which one we are.” Were we a tool or a platform? Our seat fee suggested we were a tool, while our success fee pointed to a platform. We eventually decided to start as a tool for one forwarder’s own volumes, and only bring in competing forwarders once there are enough in a region. This approach could help us use our strengths. Session 5 made an important point: “more users” alone does not create a network effect. The real value comes from better matching within a certain time and distance, which we believe only works locally. This then might make it harder to scale then we intially thought.
The second change was my mistake. I had thought neutrality was our biggest strength, since no shipping line or forwarder can be completely neutral with its competitors. But just a week before the deadline, I took a closer look at the market and found Avantida, which already offers some street turns in Europe and calls itself a neutral mediator. So neutrality went from our strongest point to something we could challenge. We shifted our focus to the forwarder side and local density. I realise now that I should have checked this in the first week and acted sooner.
The third was about risk. My first governance draft was a generic list of AI risks. The AI Canvas from Session 6 posed a narrower question: what does the AI decide within the product? Only which fields it extracts. And the errors are unbalanced. Accepting a wrong container number wastes a truck trip, while flagging a correct one costs a dispatcher a few minutes. Then Session 7 on moral hazard made me look at our own pricing. A success fee pays us for every executed turn, so we have a reason to push weak matches and inflate savings. We would, however, only push on matches that are profitable for all parties. I had been looking for risks in the technology. One of the biggest was in our business model.
What didn’t work for us and for me
Our extraction test scored 100% on 45 emails, against 84% of fields for a rule-based reader. Great, until we admitted the emails were generated and far cleaner than anything a real planner receives. If all emails were as clean as the emails in the generated product, the problem of missing the overview in these by a human would likely not exist in the first place.
We received feedback from a guy sitting next to me in class during the lecture of Session 4, he advised us to be very careful with rolling the product out in the logistics sector since they tend to be very rigid. The evidence pushed us too. By the end, our deck didn’t sell Street Turn as the surest of things. It made a conditional case: pilot with one forwarder, and scale only if the measured uplift, the savings and the guardrails hold. If these guardrails hold up, we could be looking at a product that could change the current way of operating withing the sector.
We split the work by section. Ownership was clear, but the dependencies showed up late. My failure table needed test results, and my error costs needed the economics. This meant that those cells stayed placeholders until the day of the deadline. Next time I’d agree on the shared numbers before anyone starts writing.
Working as a team, paired with AI
I used AI a lot here: to check whether I was applying course concepts correctly, to research competitors and regulation, and to help draft and restructure my sections. It made me much faster. It didn’t make me right. I still had to confirm that Avantida really works the way I described and check whether our system falls under the EU AI Act’s high-risk list for example (it doesn’t). Next to this I still had to clean up a reference list full of copy-paste leftovers at eleven at night.
AI and how I view it
That changed how I see competitive advantage. If AI makes analysis and matching cheap for everyone, it stops being a threshold and burden. This is much like what Porter said about the internet back in 2001. What stays scarce is access to data and the trust of the people whose assets you coordinate. Once these are resolved, you would still need someone who checks outputs and makes sure that these are trustworthy.
I started the course asking what AI can do more for me in assignments like this. Now I’m more interested in what becomes valuable once everyone can do it, since the threshold seems to get lower every week.
