What an agent actually does
An AI finance agent runs a loop: take a goal, gather the relevant data, reason about it, act through its tools, check the result, then either complete the task or escalate what needs a human. It is defined by the goal and the tools it's given, not by a fixed script.
Concretely, an accounts-payable agent reads an incoming invoice, matches it to the purchase order and the goods receipt (a three-way match), checks the vendor and the account coding, and then either drafts the entry for approval or flags the discrepancy for a person — rather than a human doing each of those steps by hand.
Agent vs chatbot vs RPA
A chatbot answers questions but takes no action. RPA automates a fixed sequence of clicks and breaks when a field moves or an edge case appears. An AI finance agent is given a goal and the judgement to reach it: it chooses the steps, adapts to variation, uses tools to actually do the work, and knows when to stop for a human. That's the difference between an assistant that talks and an agent that completes the task.
Governance: autonomy you can audit
Because the output affects real numbers, a finance agent has to be governable: its reasoning and tool calls are traced, guardrails constrain what it can do, and there's a clean handoff to a human for anything material. Every action is grounded in the real ledger with the supporting records linked — so any number an agent produces can be walked back to its source.