AI as a tool
This is not my first time using AI. I have used it to help write cover letters when applying for jobs, create meal plans or develop gym schedules. However, this is my first time using AI to improve a business’s processes, and I did not expect the approach to be so different. When working on personal projects, I tend to use simple prompts without paying much attention to their structure. Using AI professionally, however, requires a more strategic approach to prompt writing.
I learned that it is important to provide sufficient background information and context, clearly define the role of the AI agent, establish clear guardrails and continuously refine the output until it meets my desired outcome. Most importantly, I learned that designing an AI agent requires working backwards from the desired outcome and building the agent around that goal. These lessons showed me that AI can be a valuable tool when used carefully and strategically. This project has changed my understanding of AI and how it can be applied to business strategy.
My role and contribution
My primary responsibility in this project was the economic side of the business model. Since we chose ‘track b’, which focuses on building a technology-enabled venture, I had to develop my own cost estimates and pricing assumptions. Our project focused on helping banks prevent scams targeting senior citizens, starting in Italy before expanding to other European countries. The proposed solution consisted of a physical AI-powered device that analyses phone calls and an AI agent that provides banks with a dashboard to assess whether a transaction may be fraudulent. Since the solution involves several parts, each part required separate cost and pricing estimates.
Before estimating the costs of these parts, I researched the budgets banks, particularly Italian banks, allocate to AI. This gave a starting point for estimating the costs of the different parts of our solution. For the physical device, I used a competitor’s price as a benchmark and researched the ratio of cost of goods sold (COGS) to estimate our own hardware costs. For the AI agent, I researched cost breakdowns for AI development and used the estimated complexity of our agent to determine the likely development costs. I then developed a willingness-to-pay hypothesis based on the economic value of the fraud that our solution could prevent. This was particularly important given the upcoming EU rules expected to take effect in 2027, which could increase banks’ responsibility to reimburse customers for certain fraudulent transactions involving impersonation. Finally, I combined my research on bank budgets, development costs and willingness-to-pay with articles on B2B pricing strategies to develop a pricing model for our solution.
Important decisions and iterations
One of the most important decisions we made as a team was who would pay for our product. Initially, we considered an insurance-based model in which the senior citizen’s family would pay for the service. Under this model, we would compensate the family for financial losses resulting from false negatives, meaning scams that our system failed to detect. However, we eventually decided that the family should play a more passive role in the solution. This change followed discussions about the practical challenges. For example, we would have needed partnerships with telecommunications providers to access and analyse phone calls. Additionally, senior citizens would have had to take screenshots of suspicious messages and upload them to our app. These practical limitations led us to reconsider our business model and shift towards a B2B2C approach instead. We decided to offer our service directly to banks, including the AI-powered physical device, while allowing banks to decide whether to charge their customers for the service.
I also proposed selling the physical device at a price based on its COGS. This decision reflected the fact that hardware was not the primary value driver of our business. Our main objective was to provide a fraud-prevention service to banks. Furthermore, we wanted to eventually transition towards a software-only solution once we had established the necessary partnerships with telecommunications providers.
Beyond the economic aspects, I also contributed ideas to other parts of the project. For example, I created a slide explaining how we designed the workflow around the desired outcome and worked on a digital business model. However, neither made it into the final presentation because we concluded that they added limited value and that some of their content was repetitive.
What surprised me or did not go as planned
One of the things that surprised me most was the variety of AI agents and their different capabilities. This was my first time exploring the subject in depth, from simple reflex agents to learning agents. I learned that different types of agents require different approaches depending on their purpose. This also made estimating the development costs of an AI agent more challenging than I expected. The costs depend on several factors, including model training and maintenance. Understanding these different components helped me realise that estimating AI development costs requires more than simply calculating the initial cost of building the agent.
One aspect that did not go as planned was creating the ideal prototype to demonstrate our product. I struggled to generate a prototype that accurately reflected our vision, partly because I did not have a subscription plan. Fortunately, one of my teammates took over this task and created a prototype that clearly demonstrated the dashboard and product that we wanted to demonstrate.
How feedback and evidence changed my work
One important example of feedback influencing my work was a statistic about the number of senior citizens in Italy who had become victims of fraud. The article I first used did not appear to provide the complete picture. A teammate from Italy identified more accurate information and helped correct my interpretation. This was important because the estimate of the economic impact of fraud directly influenced my willingness-to-pay hypothesis and the pricing model.
Furthermore, discussions and evidence shared within the group influenced the overall direction of the project. We moved away from our initial B2C approach towards a B2B2C model, which made more sense when considering the practical challenges and the evolving legal responsibilities of banks.
What I learned from collaboration
The main thing I learned from collaborating with my team was the value of bringing together different perspectives and areas of expertise. We had different nationalities and cultural backgrounds, as well as different academic backgrounds, which allowed us to approach the problem from various angles. For example, one teammate studies statistics and was very skilled at analysing data. Another teammate, who is from Italy, initially proposed the idea because the problem was particularly relevant to him. He brought valuable knowledge about the local context.
I also appreciated that everyone had the opportunity to share their thoughts and that each person’s input was considered throughout the project. We discussed our ideas together and collectively decided how the final slides should look.
