Which energy contract is cheapest? That is the question every comparison website answers. When I moved to the Netherlands, even that felt impossible. The tariffs confused me, and in the end I just picked one and hoped it was a fair deal. Enerwise made me realise there is a harder question almost nobody helps with. Once you have a contract, should you switch, and is now the right time? Enerwise helps Dutch households answer it. The app calculates the real saving after any exit fee, and an AI model then explains the advice in plain language. Halfway through, a friend who was moving said they wished it existed.
In our team I took on the technical side, working on the prototype and its evaluation, and later the critical assessment and recommendation. This came naturally because of my internship at Philips, where I help build and train an AI agent. A large part of my work was developing the app in Lovable. We built it in stages, from simple input screens to the AI explanation, solar panel support, an Autopilot mode and an admin page that logs every recommendation. One principle guided the whole build. The code does the calculations, and the AI only explains them, so a wrong number from the model can never change the advice. Building it this way taught me that deciding what the AI should not do is just as important as deciding what it should do.
My other main task was the scraper. The first version of Enerwise ran on six invented tariffs, and I replaced them with real prices from supplier websites. Every supplier publishes prices differently, with some showing prices only after you enter a postcode. My first attempts failed more than once, and it took a lot of research before it finally worked. I then set up the GitHub repository and scheduled it to run weekly through GitHub Actions. I also had to decide what happens when the data is wrong, because one incorrect price would turn into confident but harmful advice. So I added checks that reject implausible prices, hold back sudden jumps, and keep the last verified price when a website fails. When suppliers had not yet published their 2027 feed-in rates for solar panels, I left those fields empty instead of filling in zero, which would have shown solar households savings that did not exist. I learned that having a working prototype means little if the data behind it cannot be trusted. I would rather show users that some information is missing than give them a confident answer based on the wrong numbers.
The biggest surprise was that the AI was the easy part and the data was the hard one. Only four of the five suppliers gave us usable prices, while each AI explanation cost less than one cent. When we ran 25 test households through the app, the same offer came out cheapest every time. This reinforced that the real question was whether switching was worth it. Although Enerwise avoided every loss-making switch, ten of the 25 AI explanations contained errors our automatic checks missed, including one invented fact. This made me question whether Autopilot was ready for real users.
Feedback shaped a lot of what we tested. In class, Prof. Li called our question, whether people would really let an agent switch for them, a strong one, but pushed all of us to make assumptions more concrete. That stuck with me, so we added a setting where users can choose between advice only and an Autopilot that switches within their own limits, with seven days to cancel. Since every household in our simulation was made up, I then showed the prototype to a few friends and asked for their honest reactions. They liked it, and half of them said they would turn on Autopilot, but fewer would actually pay for it. The others wanted to stay in control and worried about liability. A few friends are not a representative sample, but their reactions convinced me to keep advice only as the default.
Collaboration turned out to be a lesson of its own. Our team members did not all live in the same city, so we mostly worked online. It worked better than expected because we divided the project clearly, and everyone knew which part they owned. The few in-person meetings we had were especially productive. The hardest part was a side effect of that clear division. Because two of us worked almost entirely on the prototype, the rest of the team was not seeing it develop day by day, and we spent a lot of time explaining what we had built and what its limits were. If the business side did not understand that the scraper could break or that some data was missing, they could not build an honest strategy around it. Next time, I would set up short weekly demos for the whole team from the start, so explaining the prototype becomes routine instead of catch-up. My teammate and I also met almost daily because something always needed fixing, and together we recorded a demo video that works like a trailer for the app.
Close to the deadline, our team discussed how Enerwise could make money without losing people’s trust. Supplier commissions would keep the app free, but they would also give us a reason to push people to switch. I argued for a small monthly subscription for Autopilot users instead, since keeping the data up to date costs money and ongoing work, and they benefit most. I also suggested offering everyone one free analysis so people could see how it works before paying anything. Looking back, I realised we were essentially designing a freemium model, even though we had not initially approached it that way. When I helped write our final recommendation, I wanted us to be honest that the business was not yet proven. I pushed for starting small with an advice-only pilot for solar households before 2027, fixing the calculation that gave wrong advice on long contracts, and replacing scraping with data partnerships, since I had seen firsthand how fragile it was.
This project changed how I think about AI and business strategy. I started out believing the AI would be our competitive advantage. I now see it as one of the easiest parts to copy. A large comparison website could build a similar AI explanation feature, which made me think about the platform envelopment risk we discussed in class. What is much harder to copy is reliable data, permission to act for the customer, and trust, and those come from choices about partnerships, pricing, and governance rather than from the model. I see the same at my internship, where people often underestimate how long it takes to train an AI agent properly and make it trustworthy. In Enerwise, connecting the AI model took very little time. Making its advice reliable took most of the project. Competitors exist, but I believe that with better data and real user testing, Enerwise could become something households genuinely rely on, including people like my friend, and like me when I first arrived here.
