The trucks could have AI that worked most of the time. Not well enough to rely on for normal operations, but good enough for a fallback in case of losing connectivity.
Some made-up numbers to explain my point: Consider a situation where the remote operator will make the right response 99.9% of the time, failing 0.1%, and the AI is 20 times worse, making an error 2% of the time. The actual error rate will be (0.1% * connection reliability) + (2% * (1 - connection reliability)). For 3-nines reliability, which is pretty poor, the error rate will be about 0.102%. For 4-nines it will be 0.1002%.
Those sorts of figures would effectively pay for themselves in terms of safety, once you factor in the benefits for monitoring and work conditions for a desk-operator vs a physically present driver, and they don't require either a particularly reliable connection or a particularly good backup AI.