Where AI agents belong in finance, and where they do not
The rule that makes automation in finance safe enough to use every day.
Every working day, agents we built go through our clients' bank statements and supplier invoices. They code bank lines, match invoices to purchase orders and prepare drafts in Odoo. They have never posted a journal entry or paid a supplier, and they never will.
That is not a technical limitation. It is the design.
Where the time actually goes
In a growing company, the expensive part of finance is not the decisions. It is the preparation that comes before them: finding which customer a payment belongs to, checking that an invoice matches what was ordered, collecting the numbers for the month-end report. That work is repetitive, rule-based and, for a person, slow. It is exactly what an agent does well.
The decisions are different. Whether to pay a supplier this week, whether a difference on an invoice is acceptable, whether a transaction belongs to this project or another: these depend on context that lives in people's heads, and on responsibility that belongs to people.
The rule
So we split the work along that line. An agent may prepare, match, code and report. It may write into empty fields and create drafts. It may not post, approve, pay or change what a person has already entered. When it is unsure, it does not guess. It writes down what it found and asks.
In practice this means a short list every evening of the lines the agent could not place, and a set of drafts waiting for one click from the person who is responsible. The routine runs every day; the judgement stays where it was.
Why owners accept it
Business owners are rightly suspicious of anything that touches their money automatically. This rule is easy to explain and easy to check: nothing leaves the company without a human decision. Once that is clear, the conversation moves from whether to automate to what to automate next.