Anthropic released a library of agents for financial services, and two of them deserve operational attention. The month-end closer runs the closing checklist and goes as far as the journal entry. The statement auditor cross-checks balance sheets and flags inconsistencies. This isn't a demo: every tool call, every decision, every file read gets written to the Claude Console audit log, inspectable by compliance and engineering.
For anyone who has already tried to take reconciliations off the CFO's plate, the interesting point isn't the agents themselves. It's that the output isn't a black box: the auditor can sit in front of the log and walk through, step by step, what the agent did, on which data, with what reasoning. That traceability is the difference between "we automate the close" and "we automate the close without making the board of statutory auditors (collegio sindacale) scream." The real work remains, which is wiring these templates to actual data, to messy ERPs, to the exceptions every company keeps in its gut.
Why this matters for anyone building enterprise AI: the audit log as a product is the real unlock, not the agent.