Yeah that would work, but like you said it is expensive - or at least it is a cost that doesnt exist in the current model and would need specific attention. And we cannot assume customer requests / issues are uniformly distributed - it's not unlikely that 80% of issues come from 20% of customers. But if you give 80% of agents access to the same 20% of customers - you havent really solved the original problem. Albeit the risk factor is certainly lower and I would prefer this setup, who knows if that justifies the additional overhead and access control management.

Technically, the current model is a full M customer x N agents mapping. All customers served by all Agents. Adding some 0's to that mapping should certainly be possible without sacrificing much from current system. No way to know how many is reasonable

All that said, there is a natural logic that arises from this 'analysis' - Limit access to customers who are inactive, or who rarely have issues. A common sense addition would be to raise their default priority for when they do interact.