Why we keep the human in charge
It would have been easier to build a black box.
Point an AI at your accounting data, let it run, trust the output. It demos beautifully. But finance isn't a place where "usually right" is good enough: a misposted invoice isn't a quirky answer, it's a wrong number in your books that an auditor finds six months later. So from the start we made a different choice: the agent does the work, but you stay in charge. That's not a limitation we're apologising for. It's the design.
The books are not a place for guesses
A model that's 95% accurate is impressive in a demo and unacceptable in a ledger. The cost of the other 5% isn't a bad answer you shrug off: it's a restatement, a failed audit, a tax problem. So the real question was never "can the agent do this alone?" It was "who is accountable for the result?" In accounting, the answer has to be a person.
Approval is the product
Every posting the agent proposes stops for your sign-off before it touches your ledger. You see the invoice, the proposed coding, and approve or correct it. The agent does the 90% that's tedious (reading, extracting, coding, drafting), and you do the 10% that's judgment. Nothing posts without you. That single rule is what lets someone hand work to an agent and still sleep at night.
Confidence you can audit
Human-in-the-loop only works if the human can actually see what happened. So every action the agent takes is logged: what it read, what it proposed, what you approved, and when. Every number traces back to a source document and a decision. When the auditor asks "why is this here?", there's an answer, on the record. "Audit-ready" isn't a tagline for us. It's the literal output.
Autonomy is earned, not assumed
We start conservative: the agent proposes, you approve, on everything. As it proves itself on your books, you can raise the dial: let it handle the routine automatically, keep your eyes on the exceptions. Trust is granted in increments, the way you'd onboard a new hire. You don't hand the new bookkeeper the keys on day one; you check the work until you don't have to. Same here.
The point
Keeping the human in charge isn't the cautious version of an AI finance team. It's the only version that's actually trustworthy. An agent that acts without accountability isn't a finance team. It's a liability with good UX. We'd rather build the one you can put your name behind.
