Event sourcing vs CQRS: when do you actually need both?

We're building a new financial ledger service, and event sourcing feels like a natural fit for auditability and replayability. But I'm debating whether CQRS is also a necessary component or if it would be overkill for our initial use cases. We need separate read models for different views, but not necessarily in real-time immediately. When do you truly need both event sourcing and CQRS? What are the biggest challenges you've faced with event sourcing in production, especially around schema evolution for events? How do you manage projections and ensure they stay up to date and consistent without excessive complexity?