I may be missing something here, but it seems like this solves only half of the problem. Sure it's helpful to cover the earth in 3m squares with unique names, but in order to actually deliver something you need to be able to navigate to it.

The advantage of a (building number, street name, locality) address is that it identifies some point on the edge of a graph, and that the edges and vertices are well-known.

I'll grant that within any locality, people generally know how to get from point A to point B (and frequently don't know the names of the streets), but it seems like what you really need is a way to map the unique 3m squares onto that knowledge.

Even with that problem solved, another problem comes up with planning deliveries efficiently. There's no information embedded in the what3words address that encodes any notion of adjacency or relative distance. I realize that this is probably intentional to reduce the likelihood of errors in addresses going unnoticed, but it means determining nearness requires access to the whole mapping data set (or at least enough to determine that a what3words address isn't in the set of ones you care about).