In the AI Strategy Lab project we created a prototype of an AI-assisted sales process for DataVance, a data-consultancy firm which works with SMEs who do not have their own data teams. The project started with two problems that were identified during an interview with DataVance’s co-founder: consultants spend around four to six unbilled hours on each proposal and about 70% of prospects fail to convert. Our application takes the call transcript and the consultant’s notes and turns them into a one-page proposal, an illustrative sample dashboard and a cover email, all of which are approved by the consultant. In this article I consider my role, the decisions that influenced the final design, what surprised me, and the way the project has affected my thoughts on AI and strategy.
My role
Since one of my teammates works at DataVance, he took the initiative in developing the idea and building the prototype. My role focused on the strategic and communication aspects of the project. I prepared the slide deck and ensured that our design choices were grounded in the course concepts. In the early versions, we often included frameworks without translating them into concrete design decisions. I therefore started using each concept to identify a specific problem and determine how our solution could address it. One example is that Dimoka, Hong and Pavlou’s (2012) distinction between description and performance uncertainty helps to explain why prospects hesitate: they are unable to visualise the deliverable or connect its value with their own decisions. This analysis led directly to our three-part response: a sample dashboard as a signal, one real business decision as verification in the paid project, and a fixed-price starter project as a guarantee.
Key decisions and iterations
The key decision was strategic, not technical. We initially planned a two-week Power BI proof of concept on the prospect’s real data. When we tested that commitment, we saw the problem: the prospect could take the free, customised proposal and dashboard to a cheaper competitor, and using real data at this free stage raised privacy issues we had not addressed. We therefore split the offer: a free sample dashboard on illustrative data, a €2,500 fixed-scope starter project, and follow-on work once value is proven. The prospect gets essentially the same value, but capturing it is now tied to the moment the prospect commits.
A second decision concerned where the AI should and should not act. I came to see working with AI as three steps: the human specifies what is needed, the AI executes, and the human checks the output against the instructions and the prospect’s requirements. This mirrors the split between prediction and judgment/action: AI drafted follow-up questions, proposals and KPI choices, the app code calculated every displayed figure, and the consultant approved the scope, price and final email. Letting the language model generate the numbers would have been quicker, but one fictitious figure on a prospect-facing dashboard is exactly the error that damages the trust a consultancy relies on.
What surprised me
What was most surprising was the unit economics. I thought that reducing the preparation time from four to six hours to just one hour would make the business case workable. It does help, as the preparation cost per client won drops from around €1,267 to €320. Yet, with a price of €2,500 and 20 delivery hours, the starter project still only generates €280 (or 11.2%). AI has dealt with the presales cost, not the pricing. The actual lever turned out to be the offer: a price of at least €2,960 or fewer delivery hours.
I also had to come to the realisation that our evidence was limited. Even though the evaluation screen indicates 36.5 minutes of active preparation time, these figures are derived from synthetic leads and only one of the two live-demo leads was carried out. We chose to make this point clear and suggest a matched test rather than regard the figures as proof. There was a certain discomfort in writing on our own evidence slide that ‘these figures do not validate H1’, but this did make our recommendation to pilot before scaling much more credible.
Collaboration
Choosing a real company where a teammate works gave us access to experienced professionals and an interview with the co-founder, the source of our 4-6 hours and 70% figures. But it also created an information imbalance: because he understood the situation best and had direct contact with colleagues, he naturally did most of the work, and the rest of us sometimes struggled to contribute at the same level. At first, I concluded that we should have used a fictitious company, but that would have removed the very evidence that made the problem credible. The real lesson is to share context deliberately and early, for example by mapping the current sales process in week one and assigning people to different parts of the build, rather than letting whoever knows most take over. I will apply this in my own digitalisation work, where I usually have to make a process understandable to people outside the field.
How my understanding of AI and strategy changed
At the beginning of the course I mainly regarded AI as a means of carrying out existing work more quickly, I now realise that this view is incomplete. Since DataVance is able to produce a custom proposal in a matter of minutes, so too can its competitors, independent contractors and Copilot for example. Speed becomes readily available, while knowledge of the client’s source systems, the use of verified figures, accountable delivery and the client relationship remain rare. This is exactly what Barney and Reeves (2024) argue: AI on its own does not lead to a sustainable advantage, but it can enhance the resources that a firm already possesses. And in a future scenario where buyer agents compare offers, our own speed advantage vanishes as well.
The question I am now asking is not ‘what can we automate?’ but ‘what shall we still own when everybody has automated?’ Moreover, my understanding of the human in the loop has changed: at first I regarded it as a safety feature, but now I see the consultant’s judgment as being the thing that the client is actually paying for. Since my objective is to assist businesses with their digitalisation, this involves beginning with the offer and its economic aspects and considering automation as a method rather than as the end goal.
