Human-in-the-loop is a feature, not a limitation

When we tell finance leaders that every financial action in Astridex requires human approval by default, some expect us to apologize for it, as though approval were a sign the automation is not finished. It is the opposite. Human-in-the-loop is the design that makes automation safe to deploy in finance at all.

Why finance is different

A wrong recommendation in a marketing tool costs an impression. A wrong payment costs money and trust, and it touches a ledger that auditors, regulators, and a board all rely on. The cost of an error is asymmetric, so the control model has to be conservative. Approval is how you keep the asymmetry on your side.

The goal: fast, informed, rare

• Fast: an approval should take seconds because the agent has already assembled the full context.
• Informed: the approver sees what the agent did, why, and what it is asking to do next.
• Rare: routine, high-confidence work clears within guardrails, so humans approve exceptions, not everything.

Guardrails, not gates

The art is in the guardrails. You decide which actions can run autonomously within limits — say, clearing a three-way match inside tolerance — and which always need a person, like a payment above a threshold or a write-off. As trust builds, you widen the autonomous band. The approval step never disappears; it just gets pointed at the decisions that genuinely need a human.

Done well, human-in-the-loop is not friction. It is the reason a finance team can hand routine work to software and still sleep at night.