I joined this course a week before our proposal was due and joined the group only a couple of days before the deadline, which meant I had to adapt fast. It was definitely a bit challenging at the beginning, because when I joined my team they already had the idea of an AI that protects older adults in Italy from scams. So my first job was to catch up quickly.
SeniorWard
Imagine your grandmother gets a call from “you”. She gets very panicked because you need some money right now and “you” tell her not to tell anyone. SeniorWard is our answer to this which is an AI service for banks in Italy where a small device on the senior’s phone listens for signs of a scam, and if a risky transfer follows, the payment is paused and a real person at the bank checks in with the customer.
My role and contribution
Before this project, I thought scams were mostly a hacking problem. Then I saw that in Italy, about three quarters of the money lost to fraudulent transfers in the first half of 2025 was sent by the victims after a scammer talked them into it.
When we split the work, I took the parts that test the idea, asking questions like: What happens when our AI gets it wrong? And will any of this still work in a few years, when everyone, scammers included, has their own AI agent? That meant reading from Banca d’Italia’s fraud reports and Europol’s report on AI crime to the EU AI Act, GDPR, the new payment rules and Italian law on recording phone calls. I thought AI regulation was one law, but one small device turned out to fall under privacy, payment and AI rules from different angles. With that, I mapped six ways SeniorWard could fail, from a scam passing as “safe” to a senior’s savings being frozen, each with a safeguard, and tested how scammers, banks, platforms and competitors would act in 2030. My contribution was finding where SeniorWard could break and what we had to change so it wouldn’t.
How the idea changed
Our proposal at first promised a lot, with AI screening all of the calls, messages and emails, and seniors paying a monthly subscription for it. Once we worked on it more in depth we knew no bank or tech company would give a new startup access to all of those channels. So the team went small, with one country, one partner bank and phone calls only.
The next question was whether we should create an app or a physical device. An app looked smarter on paper, and our earlier discussions were more in favor of it, but I suggested my teammates really consider that mostly older people would be using it, and for many of them an app is one more thing to download and learn. A small device on the back of the phone, with a light that shows it is working, is much easier for them to trust and understand. My own decisions were mostly on the legal side. I built the analysis around Italy first and added EU law at the end. I also kept pushing us to research legal questions ourselves instead of leaving them as “needs legal advice”. The answers changed the product, and the AI can now only pause a payment while a person at the bank makes the final call.
What surprised me and what didn’t work
The biggest surprise for me was who pays. In Italy, customers carry around 91% of transfer fraud losses, not the bank, because technically they approved the payment.
My biggest struggle was the legal side. Sometimes I used an older source without realizing, and only later discovered that a newer law had replaced it. That was frustrating, because it meant redoing work I thought was finished. What saved me was going straight to the EU Artificial Intelligence Act itself, which held one more surprise. Fraud detection isn’t classed as high-risk, but reading emotions from someone’s voice is treated much more strictly. SeniorWard hence pays attention to what is said rather than how it sounds.
Evidence and feedback
Monetization was a big discussion and our team had nearly made SeniorWard a public service funded by the government. Then feedback within the team changed to who pays. One teammate suggested refunding seniors for their losses, and another suggested that the bank should pay instead. We went with the second idea, because new EU rules will make banks carry more of the fraud losses.
As for evidence, the biggest realization for me was finding out that Google already offers free scam detection on its newer Pixel phones in Italy. Phone platforms see the call but not the payment, and banks see the payment but not the call. Since SeniorWard is the only one that links the two, my future-market test found that it survives as a service for banks rather than a device. Additionally, in our scripted testing, SeniorWard paused four of the five scams while the bank’s regular checks paused none.
Working as a team
Since I joined late I had to ask a lot of questions but also not too much that my teammates did not think I trusted their decisions and still find a way to add value.
Every part of the assignment was way more integrated compared to a regular assignment. For example, my failure analysis depended on how the prototype worked, and my future test depended on the business model my teammates were building, so we had to stay in close contact. When we built the prototype, the device, the call analysis, the bank’s dashboard and the test scenarios all had to tell the same story. So it was a lot of messages and calls.
I also saw how much an international team created value. The idea to start in Italy came from our Italian teammate, who knew the market and the culture better than us.
Main Takeaways
First, with all the talk about agentic AI in our lectures, I honestly thought AI meant humans could relax and take it easy. However, the biggest thing I learned with SeniorWard is that every mistake has to land on a real person so i learned that the smarter AI gets, the more it matters who is governing it.
Second, I learned that the AI itself is rarely what makes a business strong. As someone who uses the AI buzzword a lot because it’s literally everywhere, I really did think putting AI anywhere would make it work better. Seeing Google already offer free scam detection made me realize a startup can’t win just by detecting scams, and it made what we talked about in the lecture click in real time.
Third, I learned that in AI, information gets old very fast (I knew it theoretically) but I really saw it during this assignment a lot. The AI Act was changed this year, a fine I wanted to use had already been cancelled, and our business case relies on a payment rule that only starts in 2027. So checking matters as much as finding.
