How to prepare variance commentary for management accounts with an AI agent

Split variance commentary into what moved and why. The agent computes variances against budget and forecast, finds the transactions that drove each one, drafts the factual part and asks the budget holder a specific question about the cause. Budget holders answer, FP&A edits the commentary into a view of the business, and the finance director signs the management accounts.

What is variance commentary, and who owns it?

Management accounts compare actual results with budget, forecast and prior periods. Commentary explains the differences that matter, so managers can act. Good commentary says what drove a variance, whether it will continue and what is being done. Weak commentary restates the numbers in words.

FP&A or business partners own the commentary. Budget holders own the reasons in their areas. The finance director signs the management accounts. We argue that variance commentary needs a person's judgement: the agent can find the facts, but whether a variance is a timing blip or a trend is a business view. Whether an agent can prepare management accounts more broadly follows the same split.

What are the steps, and who acts at each one?

1. Agent, depth 1: computes variances. Actual against budget, latest forecast and prior year, by line and cost centre, at the level of the chart of accounts the pack uses.
2. Agent, depth 1: applies materiality. Thresholds set by FP&A, by line and in total. Only material variances get commentary.
3. Agent, depth 1: finds drivers. Breaks each material variance into its parts: volume against price, timing against permanent, and the specific transactions. For cost lines, it looks at postings from accounts payable, payroll, accruals and journals.
4. Agent, depth 2: drafts the factual line. "Marketing is 60,000 over budget: 45,000 from an agency invoice for a campaign not in budget, 15,000 from a timing difference on events moved from next quarter."
5. Agent, depth 2: asks the budget holder. A specific question, not "please explain": "Is the agency campaign a one-off, or will further invoices follow?" Under certification, these requests can be sent at depth 3.
6. Person, budget holder: answers. Cause, outlook and action.
7. Person, FP&A: writes the commentary. Combines facts and answers into a view, removes what does not matter, adds what the numbers do not show.
8. Person, finance director: signs the management accounts.

What separates a useful line from a useless one?

A useless line: "Staff costs are 90,000 over budget due to higher staff costs." It restates the number. A factual line from the agent: "Staff costs are 90,000 over budget: 55,000 from six contractors in engineering not in budget, 35,000 from overtime in the warehouse." A useful line after the budget holder answers and FP&A edits: "Staff costs are 90,000 over budget. Engineering is covering two vacant roles with contractors until hires start next quarter; warehouse overtime reflects a seasonal peak that will ease. Expect the variance to halve next month."

Which controls matter, and what evidence do they leave?

References on every factual line. The postings or invoices behind each number are linked. Evidence: references stored with the draft.

Causes come from people. The agent does not write a cause it did not receive from a budget holder or find stated in a document. Evidence: the budget holder's answer linked to the line.

Materiality version. Evidence: thresholds used, approved by FP&A.

Commentary sign-off. Evidence: the finance director's sign-off on the final version.

Consistency with the flux review. Where the close flux analysis already explained an account, the commentary should agree with it. Evidence: cross-references.

What should you measure?

• Answers received from budget holders before the deadline.
• Factual lines changed by FP&A, a sign of draft quality.
• Commentary that turned out wrong the following month.
• Days from close to signed management accounts.
• Repeated variances explained as "timing" for more than two months.

Where does it usually go wrong?

Mechanical commentary. The pack is full of correct but useless lines. FP&A must cut, not just accept, and lead with what matters most.

Timing that never reverses. A variance is called timing for three months running. The agent should track variances called timing and flag those that have not reversed.

Budget phasing errors. The budget was spread evenly across months while spending is seasonal, so variances appear every month and mean nothing. Fix the phasing; the agent can show which lines have phasing patterns that do not match history.

Invented causes. An agent fills gaps with plausible reasons. Requiring a source for every cause stops this.

Questions ignored. Budget holders do not answer the agent. Make answers part of their close duties and report non-response to the finance director.

Where Pryme fits

In Pryme, computing variances, tracing drivers and asking budget holders are separate capabilities, certified on your own past management accounts. Postings arrive sealed and budget holders' answers declared, so the finance director can see which part of each line is data and which is someone's account. Every draft, answer and edit is filed. The finance bounds are in AI agents for finance.