I was surprised to learn that an inertial navigation system even works at all; I'd expect that the integration drift would throw the velocity measurement off badly enough, not to mention the position. Other questions have asked about how these systems work, and the answers have explained that cross-checks with other measurements which give absolute position (e.g. GPS) improve accuracy.
But given a state-of-the art INS which operates in isolation (no interfacing with airspeed, GPS, etc) from takeoff, how much drift could we expect for, say, a normal transatlantic flight? Does this change if the aircraft flies through turbulence?