I understand what you're saying, but in the example for accounting, we can also solve this problem using NoSQL. Because the most important feature we're talking about there is transaction support. Similarly, schema-on-write can be provided by a library.
To me it seems like NoSQL works better when there is less to normalize, which is the case with microservices. Those services struggle with the support for a distributed transaction when they have to make a distributed transaction. This problem will be very easily solved in SQL (assuming its not shared to completely denormalize everything for performance).
Note that this normalization problem also shows up in schema-on-write. If multiple people are contributing to a schema from different teams, then it will become hard to maintain a schema-on-write.
To me it seems like NoSQL works better when there is less to normalize, which is the case with microservices. Those services struggle with the support for a distributed transaction when they have to make a distributed transaction. This problem will be very easily solved in SQL (assuming its not shared to completely denormalize everything for performance).
Note that this normalization problem also shows up in schema-on-write. If multiple people are contributing to a schema from different teams, then it will become hard to maintain a schema-on-write.