The gap between business demands and data integrity
AI adoption in enterprises is now widespread, but according to Loris Lombardo, product manager at Horsa Insight, speaking at the Innovation Challenge 2026 conference—Solutions to avoid disillusionment with AI, promoted by the Digital Innovation Observatories of Politecnico di Milano—the requests from operational departments are very concrete: support on investments, month-end closing, and financial and economic reporting. These are practical needs that virtual assistants should meet quickly.
The problem, Lombardo explained, emerges on the technical side: behind a simple interface lies complex analytical processing, and when information sources are inconsistent, responses are inaccurate or incomplete. The risk, he warned, is that trust in the tools erodes to the point of disillusionment, dampening initial enthusiasm.
The decision as a product, with humans in charge
Horsa Insight’s proposed response is a decision-driven approach inspired by the Politecnico Observatory’s definition: managing, designing, and executing decisions as if they were real products. In this setup, knowledge is structured before algorithmic processing, defining accuracy requirements and human responsibilities in advance.
Lombardo specified that the tools aim to augment and support decisions, leaving the final choice to the user. Technology therefore does not replace the managerial component but strengthens its evaluative capacity, maintaining human governance over the systems’ output.
Explorers and decision-makers: two different needs
The analysis distinguishes two categories of users with opposite error tolerances. Explorers, or power users, have technical backgrounds and conduct in-depth analytical sessions: they can accept margins of approximation because they are able to critically evaluate the output. Decision-makers—executives and department heads—must deliberate within tight timeframes and have neither the time nor the expertise to verify the calculation process: for them, an inaccuracy compromises the validity of the choice.
The cited case concerns a global fashion brand active in e-commerce and digital marketing, with data from multiple advertising platforms collected in separate silos, inconsistent KPIs, and slowed decisions. The initiative is based on two components: a Semantic layer that unifies calculation rules, KPIs, and attribution models in data products, and AI agents specialized by role. The available material cuts off right in the description of the second component, without indicating measurable results for the case.
The architecture’s actual effectiveness, the implementation figures, and the details of how the agents work therefore remain to be verified—elements that the provided text does not document.
