On client side I would use vector layers, because they are more versatile than raster layers.
One of the biggest issues of vector layers is that the client can collapse when the number of shown features is too big. Fortunately, OpenLayers has tools to cope with problems like this, like the cluster strategy.
On the server side, I would use shapefiles if I was the only person doing the backoffice work that you mentioned; in this scenario there would be no need to set up an spatial database like PostGIS, a non-trivial task (KISS principle). Another spatial database that you can find interesting is SpatialLite.
In case internationalization is a must, shapefiles might not be a suitable format. Should this be the case, you can try ArcGIS' file geodatabase (FGBD). It seems that Quantum GIS supports FGDB via GDAL.
EDITION: I complete my answer after some criticism in some of the comments.
While vector layers might consume more bandwidth than raster layers, this issue is addressable. First, in case the vector data is retrieved using WFS, there is no obligation to use GML; the client can get the data in the less verbose GeoJSON format. On the other hand, the vector data returned by the WFS server usually includes attributes, which are not needed to render the feature. Any decent WFS server can be configured to not send such attributes, thus getting savings in bandwidth usage.
Another mentioned issue was that vector layers produce maps that are slow to render. In regard to rendering, the difference between raster and vector layers is where this task is done: raster layers are rendered by the server and vector layers are rendered by the client. While old web browsers could not be very good at rendering vector layers, the current version of the main ones - Chrome, Firefox and Internet Explorer... - do it much better because the improvements introduced in their JavaScript engines. On the mobile arena, the computing power of these devices have dramatically improved during last years; in fact, is not unsual that smartphones have multicore processors.
About the issue of an too big number of features on client side, it can be said that WFS allows limiting the maximum number of returned features. In case features are made of a too big number of points, Douglas-Peucker simplification can be calculated.
And last, about the superior versatility of the vector layers I mentioned, in the section of OpenLayers examples there are some examples that users can find useful, like resizing, rotating or selecting features, to mention some.