Management accounts: a sensible first job for an AI agent
The monthly management accounts pack is the same report, built the same way, every month. Why it suits an agent that drafts while a person approves.

Every month, somewhere around working day five, a finance team produces the management accounts pack. Profit and loss against budget, balance sheet, cash, a few KPIs, a page of commentary. The numbers change. The pack does not. Same layout, same sources, same checks, same reader, every month. That makes the management accounts pack one of the most sensible first jobs for an AI agent: not because it is unimportant, but because it is recurring report production with a clear shape, a clear reviewer and a clear definition of done.
Why the pack is report production, not analysis
Watch the management accounts being prepared and the job splits into two very different kinds of work.
The first kind is production. Run the trial balance once the close is signed off. Map it into the pack layout. Pull the budget and last year's figures. Refresh the charts. Calculate the KPIs. Check that the balance sheet balances and that the profit in the pack agrees to the ledger. Flag every variance over the threshold. This is most of the hours, and it is rule bound from start to finish.
The second kind is judgement. Decide which variances matter. Explain why margin fell. Say what the board should worry about and what it should not. This is the part the reader actually pays for, and it is usually squeezed into the last afternoon because production took the rest of the week.
Most teams treat both as one task. Split them, and the production half looks like what it is: a recurring task with steps that can be written down.
What makes the management accounts pack a good first job
In an earlier post we set out the ladder from drafts to trust: an agent drafts and a person approves, then the agent acts and is reviewed, then it acts alone, earned per task. Not every task suits the first rung. The pack does, for five reasons.
- It repeats on a fixed calendar. Monthly, on a known working day, with the same inputs. Every month's draft can be compared with the last.
- The output is checkable. The pack has to tie back to the ledger. Totals agree or they do not. A reviewer can see in minutes whether the draft is right.
- Nothing leaves the building until a person signs it. The blast radius of a wrong draft is a correction, not a payment that cannot be recalled or a filing that cannot be withdrawn.
- The reviewer already exists. Someone senior reads the pack before it goes out today. The approval step is not new work. It is the step you already have.
- The method is usually half written already. The pack template, the mapping and last month's version are a large part of the instructions. What is missing tends to be the thresholds and the judgement rules, which need writing down anyway.
What the agent drafts, and what stays with people
Here is a sensible split for a hypothetical mid-size finance team. The detail will differ in yours.
The agent drafts
- The mapped P&L, balance sheet and cash flow, from the signed-off trial balance.
- Budget and prior year comparisons, with variances calculated.
- A list of every variance over the written threshold, each with the ledger lines behind it.
- The KPI page, recalculated from defined formulas.
- A first pass at factual commentary: "Payroll costs were above budget by the amount shown, mainly in the two lines listed."
- A checklist showing which tie-outs passed and which did not.
A person decides
- Whether the close is clean enough to start the pack at all.
- Which variances matter and why, and what the commentary should say about them.
- Anything that involves a judgement call on accruals, provisions or one-off items.
- Whether the pack is ready to send, and to whom.
The financial controller still signs the pack. They just start from a reconciled draft on day five instead of a blank template on day three.
What to write down before the agent starts
An agent runs on the written method, so the gaps that a person fills from memory have to be closed first. Before the first draft, write down:
- Inputs. Which report, from which system, run at what point in the close, and what to do if it is not ready.
- Mapping. Which nominal codes feed each line of the pack, including the ones someone moved by hand last year and never documented.
- Thresholds. What counts as a variance worth flagging: a percentage, an amount, or both.
- Tie-outs. The checks that must pass before the draft goes to review, written as a list.
- Stop rules. When the agent should stop and ask rather than carry on: a missing input, a code it has not seen before, a check that fails.
- Definition of done. What "draft ready for review" means, so nobody argues about it on day five.
The same list lets someone cover the pack when the usual preparer is on holiday.
How the reviewer's corrections teach the agent
The first few months of drafts will need corrections. That is expected. The value comes from capturing them instead of quietly fixing them. Log each one with what was wrong and why: a mapping gap, a bad input, a judgement the agent should not have made, or a plain error. Then change the written method, not just the pack. We covered this in more depth in designing the correction loop.
After, say, six months of drafts, the log tells you something real. If the reviewer is now changing commentary wording and nothing else, the production half has earned more freedom. If they are still fixing mappings, it has not, and the fix is in the instructions.
What it means for the people who build the pack today
Be plain with the team about this. The people who assemble the pack know where it breaks: which code always lands in the wrong line, which budget file is never final in time. Ask them to write the method, and say what it is for. The production half of the pack is the dull part of a skilled job. Taking it away first gives the preparer the time to do the part the board actually reads, and that is the part that builds a career.
One thing to do this week
Take last month's management accounts pack and list every step it took to produce, from the trial balance to the email that sent it. Mark each step as production or judgement. Count the hours on each side. That one page is the starting point for writing the method down, whether or not an agent ever reads it.


