Two things, and only two: the quality of what we learn at the first meeting, and the speed at which we make corrections when the system is wrong. The rest of the work can be done very well remotely, and we say so instead of selling proximity as a value in itself.
The first point is the most concrete. You understand a repetitive task by watching someone perform it: the three tabs they open without thinking, the parallel spreadsheet no one mentioned in the meeting, the reason a certain inquiry is always set aside. These details determine whether an agent is useful, and they are easier to see while sitting next to the workstation than on a video call.
The second point concerns trial operation. When the agent proposes something incorrect, the person who noticed it needs thirty seconds to show it, not a written report. At a site in Île-de-France, those thirty seconds happen in person.
Outside those two moments, we work remotely as we do everywhere else in France: we do not multiply meetings for convenience.