Small wheeled delivery robots - roughly the size of a large cool box, moving at walking pace along pavements, carrying a meal or a parcel - have moved from demonstration to routine operation in a number of cities. They are unglamorous and they are slow, and both of those are the point. They work in a narrow set of conditions: short distances, light loads, dense areas where a great many destinations sit within a couple of kilometres of the collection point, and pavements that are wide, flat and in good repair. Outside those conditions they are not a proposition at all, which is why they appear in some places and not others without much apparent logic.
The technically interesting problem turned out not to be navigation. Getting a machine to find its way along a pavement and around a parked car is largely solved. The difficulty is people - not as obstacles, since avoiding a pedestrian is straightforward, but as participants. People step in front of the machine to look at it. Children follow it. Somebody will block it to see what it does. A few will try to tip it over or take what is inside. The machine has to keep going through all of this without either stopping permanently or pushing through, and it has to do so in a way that does not make the pavement worse for everybody else using it.
That is why most of these systems keep a person in the loop remotely, one operator watching several machines and taking over when one gets stuck. This is often reported as a gap between the promise and the reality, as though remote assistance were an admission of failure. It is better understood as the actual design. The machine handles the ordinary case, which is nearly all of it, and a human handles the exceptions, which are rare but numerous in kind and will not be enumerated in advance. That arrangement - autonomy for the routine, a person for the unexpected - is how most working automation is actually structured, and it is much more common than full autonomy.
For a small business the direct relevance is limited: these are not machines anyone is going to buy, and the economics depend on running many of them from one base in a dense area. The transferable observation is about where the hard part sits. The engineering problem was solved before the social one, and the social one - what happens when a machine operates among people who did not agree to it and have their own ideas about it - is what determines whether it works in a given street. That is worth remembering with any automation that leaves the back office and meets the public, including the far less dramatic kinds that a small business might actually adopt.