Drafts first, trust later: how to bring AI into a finance team

The usual worry about AI in a finance team is not whether the technology works. It is what happens the first time it gets something wrong. A miscoded invoice, a reconciliation marked as done when it is not, a client report with last month's number in it. In finance, one confident mistake can undo a year of goodwill. That worry is healthy, and the answer is not a bigger pilot. It is a ladder: the agent drafts, a person approves, and freedom is earned one task at a time.

Bringing AI into a finance team: trust is earned per task

Think about how a new junior joins a firm. Nobody hands them the year end pack in week one. They start with simple bank reconciliations, a senior checks every one, and after a few clean months the checks get lighter. The same person might be trusted completely with AP runs and still be learning the VAT return.

That is the model worth copying for an AI agent. Trust is not a switch you flip for "AI" as a whole. It is a setting per task, and it moves only when there is evidence to move it.

The opposite approach, switching a tool on across a whole process and hoping, tends to fail everywhere at once. The team then concludes AI does not work for them, when what failed was the rollout.

The three rungs of the ladder

There are three levels of freedom an agent can have on a recurring task. Each task starts at the bottom.

Rung one: the agent drafts, a person approves

The agent prepares the work and stops. It matches the bank lines, proposes the journal, fills in the report template, drafts the client email. A named person reviews the draft and either approves it, corrects it or rejects it. Nothing leaves the building, and nothing posts to the ledger, without that approval.

This rung is where most of the learning happens. The reviewer sees where the agent is reliable and where it guesses, and the corrections show where the task's instructions are thin. Often the most useful outcome of the first month is the list of things nobody had ever written down.

Rung two: the agent acts, a person reviews after

Once a task has a run of clean drafts, the agent can complete it and a person reviews the result afterwards, in a batch or on a sample. The agent posts the routine bank matches itself and sends anything unusual to a person. The reviewer looks at the exceptions and a sample of the rest, rather than every line.

The difference from rung one is timing. The work is done before the review, so the agent's mistakes need to be cheap to find and cheap to fix. Tasks where a mistake is quietly reversible fit here. Tasks where a mistake reaches a client or a regulator before anyone sees it usually do not.

Rung three: the agent acts alone

Some tasks will earn full autonomy: the agent runs them, logs what it did, and only raises a flag when something falls outside the rules. A person still owns the task in the sense that matters. They are accountable for it, they look at the log, and they decide if it should move back down a rung.

Plenty of tasks should never reach this rung, and that is fine. A routine payroll check and a judgement call on a bad debt provision are different kinds of work. The ladder makes that difference visible; it is not there to push everything to the top.

What earns a promotion up the ladder

If trust is per task, you need a way to decide when a task moves. Gut feel from the reviewer is a start, but it is not a record. Three things are worth tracking for every task an agent touches:

• Approval rate. Out of the drafts it produced, how many were approved without changes.
• Correction type. A changed description is cosmetic. A wrong account code or a missed accrual is substantive. Track which kind they are.
• Exceptions returned. How often the agent correctly said "this one is not for me" and handed back to a person. An agent that never returns anything is more worrying than one that returns too much.

Set a simple rule in advance. As an illustration, a task moves from rung one to rung two after three consecutive months with no substantive corrections, and moves back down after any substantive error. The exact numbers matter less than writing them down before you start, so promotion is a decision based on evidence, not on enthusiasm.

Choosing the first tasks

The first tasks on the ladder should run often, follow clear rules, be written down and be cheap to get wrong. Frequency builds a track record quickly, documentation is what the agent works from, and a low blast radius means early mistakes cost a correction, not a client.

Bank reconciliations for straightforward accounts, supplier statement matching and recurring report preparation tend to fit. Accruals with a lot of judgement, anything involving a client's tax position, and tasks only one person knows tend not to, at least not yet. If a task lives in one person's head, it is not ready for anyone else, human or agent. Our post on key person risk in the month end close covers how to get that know-how out first.

All of this assumes you know which recurring tasks your team runs in the first place. If you do not, start there; most firms have no list of what their team does, and the ladder needs one.

Tell the team how the ladder works

The people reviewing the drafts are the people who do the work today. They should know exactly which tasks are on the ladder, what rung each is on, what the promotion rule is, and what their role becomes as tasks move up. Hiding any of it is wrong and self-defeating, because the reviews are only as honest as the trust behind them.

Framed properly, this is good news for them. The first tasks to move are the repetitive ones that fill the close week and stop people doing the work they trained for. The reviewer becomes the person who decides when the agent is good enough, which is a more senior job than the one they had. That is how Beaks is designed to treat agents: as a teammate with its own workload and track record, starting with drafts, with every handover logged.

One thing to do this week

Pick one recurring task your team runs at least weekly, with clear rules and a written process. Write down, on one page, what rung one would look like for it: what the agent would draft, who would approve it, and the rule that would move it to rung two. You do not need an agent to do this. The page itself will tell you whether the task is ready, and if it is not, exactly what is missing.