Traditional RPA scripts a fixed sequence of clicks against a specific screen layout. It's fast and cheap on the exact happy path it was built for, and it breaks the moment a field moves, a login page changes, or a transaction falls even slightly outside the scripted flow. Someone then has to notice the failure and finish the work by hand anyway, usually after the fact, during close week, when there's the least slack to absorb it.
Finance work is full of exactly the cases RPA struggles with: the bank transaction that's a timing difference rather than a real discrepancy, the accrual that needs to deviate from last month's pattern because the underlying contract changed, the vendor who needs a firmer tone on the third follow-up than the first. That's most of the actual volume in a reconciliation or AP queue, not a rare edge case, which is exactly why RPA deployments in finance so often end up needing as much manual babysitting as the process they replaced.
An agent doesn't run a fixed script. It reads the transaction, the account history, and the relevant policy, reasons about what's actually going on, and takes the action it's confident in. When it isn't confident, it stops and asks, with its reasoning attached, instead of erroring out silently and waiting for someone to notice three days later that a queue has been piling up.