Use AI to run a workflow without losing control: a human-in-the-loop playbook

How lean finance teams get the speed of automation without the "what did the bot just do?" risk.

The fear with AI in finance is reasonable: hand a process to an autonomous agent and you're one confident mistake away from the wrong thing going to your best customer. But the choice isn't "do it all by hand" or "let the robot run free." The teams getting real leverage from AI use a human-in-the-loop pattern — the AI does the work, a person approves it, and trusted steps graduate to automatic over time.

Here's the playbook, drawn from running it on a real recurring workflow.

What AI is genuinely good at (and what to keep off it)

AI is excellent at the thinking parts of a workflow: researching, classifying, drafting, summarising, scoring, and scheduling. It is not a fire-and-forget cannon — especially for anything that touches a customer or moves money. So split the work:

Let AI draft and prepare. Classify the inbound, draft the reply, pull the data, propose the next step.
Keep a human on the decision. Review and approve before anything sends or posts.

The five moves that make it run

Promote your rules into a reusable "skill." Your tone, your do's and don'ts, your best examples, your hard rules — write them down once so every draft follows them. This is the single highest-leverage step: it's what stops quality from drifting as volume grows.
Use scheduled tasks for the recurring parts. A daily "pull what came in, classify it, and draft responses for my review" run turns an ad-hoc chore into a system.
Keep a surgical-edit loop. Don't re-run the whole batch to fix one output — tell the AI "make this less formal" or "add the due date" and let it adjust just that one.
Let connectors close the loop. Once something is approved, have it written back into your system of record automatically, so the audit trail stays clean without manual entry.
Graduate to autonomy on your terms. Start with every output reviewed. As a category proves itself, let the AI send those on its own — one classification at a time, never all at once.

Why templates make automation safe

The reason a step becomes safe to automate is rarely "the AI got smarter." It's that the output is constrained. When a reply is drafted from a template whose blanks are filled with live data from your system — invoice number, amount, due date — there's very little room for it to go wrong. Templated, data-bound responses are both more accurate and the natural candidates to let run automatically first.

Map it to finance

The same pattern applies directly to AR and finance ops:

Inbound debtor replies — classify (promise to pay, dispute, query), draft from a template, approve, then auto-send the safe categories.
Reminder drafting — generate personalised chasers; approve the batch.
Reconciliation prep — let AI match and flag exceptions; a human clears the edge cases.

The honest takeaway

You don't need to trust AI completely to benefit from it. You need a process where it does the heavy lifting, you stay in control of what matters, and automation earns its way in. That's not a compromise — it's the version that actually survives contact with real customers.