For our AI Strategy Lab project, we developed a prototype of an AI-assisted sales process for the company I work for. The idea was to solve two challenges at once. First, we wanted to automate much of the time-consuming process of preparing proposals after the first client meeting, reducing the hours consultants spend on potential customers and thereby saving money. Second, we wanted to make the proposed solution more tangible for customers. Instead of receiving only a written proposal, they would also receive an interactive demo dashboard with illustrative data, allowing them to visualize what the final solution could look like. Our hypothesis was that this would make it easier for customers to understand the value of the service and hopefully lead to a higher conversion rate.
The final prototype follows four main stages: processing a client meeting transcript, generating a tailored proposal, creating an illustrative dashboard, and preparing the documents and email for the customer. Throughout this process, the consultant remains involved to review the AI’s output and make adjustments where necessary.
My contributions
My main contribution started with identifying the opportunity. Since I worked at the company, I took responsibility for speaking with my colleagues and conducting interviews to find out where AI could make a meaningful difference.
After our team agreed on the concept, I took responsibility for turning it into a working prototype. I started developing the application in VS Code, connecting it to an AI model through an API and running everything locally. One of my first designs divided the process into nine separate steps. On paper, this seemed logical because every activity had its own place. However, once I started looking at the application from a user’s perspective, I realized that navigating nine steps was neither intuitive nor visually appealing. I eventually redesigned the interface around four main stages, each containing several smaller substeps. This made the process easier to follow without losing the functionality we needed.
Another challenge was deciding which features were worth pursuing. Initially, I wanted the application to send the proposal and accompanying email directly to the client. I spent considerable time trying to implement this, but I could not get it working reliably. Eventually, I decided to change direction. Instead of automatically sending the email, the application would generate a message that could easily be copied, while the proposal and dashboard would be available as downloadable files. At first, removing a feature I had spent so much time developing felt like a compromise. Looking back, it was probably the better decision. It simplified the application while maintaining most of the value for its users.
The importance of keeping humans involved
One principle I cared about from the beginning was keeping a human in the loop. I did not want consultants to blindly accept whatever the AI generated, particularly when the output would eventually be shared with potential clients. For example, consultants needed to edit the generated proposal directly within the application. We also discovered that AI-generated follow-up questions were not always sufficient. Sometimes a consultant knows something important that the AI simply does not think to ask. Therefore, we added the possibility of providing additional human input independently.
Another interesting piece of feedback was that the proposal should be no longer than one page. While AI can easily produce lengthy and detailed documents, this was not necessarily what the consultants needed or what they believed would be most useful to potential customers.
These changes showed me how important it is to test assumptions with real users. I had focused heavily on making the technology function correctly, but testing reminded me that the people using the application ultimately determine whether it is useful.
A working demo is not a finished product
Because we connected the application to DeepSeek through an API, we had to think carefully about what information would be shared with an external AI provider, how that information might be processed, and whether using sensitive company or client information would be acceptable. These questions have not all been resolved. The application currently works as a demonstration, but it is not yet ready for unrestricted everyday use. Security, privacy, and reliability requirements still need further work.
Our testing showed promising results, with the AI-assisted workflow taking roughly 10–15 minutes in our test runs. However, this is not yet a controlled comparison with the existing manual process. More importantly, we have not proven that the new approach actually improves conversion rates.
Working with a team outside my company
Because I was familiar with the company, I sometimes found it difficult to explain the business context to my teammates. Certain processes and decisions made sense to me because I encountered them in my work, but they were not necessarily obvious to someone looking at the company from the outside. Our collaboration went well overall, and we organized both online and physical meetings to discuss the project, share progress, and work on the final presentation and research. Nevertheless, I underestimated how much effort was required to communicate the context to my team members.
If I could restart the project, I would seriously consider choosing a fictional company instead. That would give every team member a more equal starting point and might make collaboration easier. Of course, this would also mean losing one of the project’s greatest strengths: the opportunity to interview actual employees, receive practical feedback, and potentially implement the result. This is a trade-off I had not fully considered at the beginning.
How my understanding of AI strategy changed
Initially, I mainly saw AI as a tool for making existing tasks faster. While I still believe that automation can create a lot of value, I now understand that its success depends on how it is integrated into a business process. This connects to Barney and Reeves’ argument that AI alone is unlikely to create a sustainable competitive advantage. The greater opportunity lies in combining it with capabilities and expertise that are specific to a company.
For our project, the real opportunity was not simply generating proposals automatically. It was combining AI with industry knowledge, consultant judgment, and a more accessible offering for potential customers. Perhaps the most important distinction I now make is between technical feasibility and strategic value. We demonstrated that our application could work, but we have not yet proven that it will increase customer conversion or deliver the intended business benefits.
This project is not yet finished
My next step is to hand over the prototype to my employer and continue improving it, particularly the security, reliability, and remaining practical requirements. The goal is to eventually turn the demonstration into an application that can genuinely support the sales process. If there is one lesson I will take into future projects, it is that successful AI implementation is not about automating as much as possible. It is about knowing what should be automated, what should remain in human hands, and how the combination creates real value.
