Your boss may not know much about hard graph problems, but are you so sure this is impossible? For one thing, many problems are not as general as full TSP. Are you sure your problem doesn't reduce to finding a Hamiltonian cycle, for example? Perhaps your graph has geometric features which can reduce the complexity? For example, does it satisfy the triangle inequality?
Regardless, for most real applications coming close to the optimal solution (as opposed to solving exactly) is perfectly acceptable. In the case of traveling salesman and related problems, there are many heuristics which can achieve this. Have you considered and rejected branch and bound, cutting plane methods, minimum spanning trees, simulated annealing, or Kernighan-Lin heuristics?
In my experience, real-life problems often have weaknesses that can be exploited from an algorithmic perspective. It's mostly just a question of modelling your problem correctly.