Using AI to prepare management reports.

Goal: Produce a draft of this month’s management commentary in which every figure ties to your accounts and every explanation comes from a named person. Who it is for: Finance managers, business owners, practice accountants and operations leads who write the monthly report for owners or a board. Time: 40 minutes, plus last month’s profit and loss against budget and a few lines of notes from the managers responsible.

“Revenue increased due to strong seasonal demand.” Any reader of management reports has seen a sentence like that, and an AI will write one for every variance you give it. It sounds informed. It is a guess. Management commentary is valuable precisely because someone who knows the business explains what happened, so the method here splits the job: the assistant does the arithmetic, the layout and the plain wording, and people supply the reasons.

What AI does well and badly in a management report

AI is strong at variance arithmetic, consistent formatting and turning terse notes into readable sentences. It is weak at causes, materiality judgements and knowing which numbers the board is sensitive about.

Exercise: draft this month’s commentary from real figures

1. Export a summary P&L. From Xero, MYOB or your system, export actual versus budget (or prior year) for the month, excluding GST, condensed to 10 to 20 lines.
2. Collect one-line notes. Ask each manager for a sentence on anything unusual in their area. Write their name beside each note.
3. Set materiality. Choose a rule such as “comment on variances over 5% and over $2,000”. Smaller businesses may use lower dollar figures.
4. Run the drafting prompt. Paste the figures and notes into ChatGPT, Claude, Copilot or Gemini using the worked example prompt below.
5. Tie it out. Check that totals in the draft equal your P&L, recompute two variances by hand, and check each favourable or unfavourable label.
6. Chase the placeholders. Send every [Reason not supplied] line to the responsible manager and replace it with their answer, or state plainly that the cause is still being investigated.
7. Final read by the person who signs. They own the commentary, not the assistant.

Worked example: a building supplies retailer’s September

Illustrative example: a 22-person Hamilton building supplies retailer preparing its September 2026 report. The business and figures are fictional.

Draft management commentary for September 2026 from the data below.
Amounts are NZD excluding GST.

Line, Actual, Budget
Trade sales, 412000, 380000
Retail sales, 96500, 110000
Cost of goods sold, 309800, 294000
Wages, 88200, 86000
Freight, 14900, 11000
Rent, 18500, 18500

Manager notes:
- Trade: two large orders from a subdivision builder. (Jenny)
- Freight: carrier fuel surcharge increased from 1 September. (Mark)

Rules:
1. Show a table of material variances: over 5% AND over $2,000.
   Give $ and %, and mark F (favourable) or U (unfavourable). Higher
   costs are unfavourable.
2. Calculate total sales and gross margin % for actual and budget.
3. Explain variances ONLY using the notes, naming the source. If a
   material variance has no note, write [Reason not supplied].
4. Any suggestion must be phrased as a question to check.
5. Plain New Zealand English. No adjectives like "strong" or "soft".

Example output (illustrative, abridged):

Material variances, September 2026 (excl GST)
Line                Actual     Budget     Variance
Trade sales         $412,000   $380,000   +$32,000 (8.4%) F
Retail sales        $96,500    $110,000   -$13,500 (12.3%) U
Cost of goods sold  $309,800   $294,000   +$15,800 (5.4%) U
Freight             $14,900    $11,000    +$3,900 (35.5%) U

Total sales were $508,500, $18,500 (3.8%) above budget. Trade sales were
$32,000 over budget, from two large orders placed by a subdivision
builder (Jenny). Retail sales were $13,500 under budget.
[Reason not supplied]
Gross margin was 39.1% against a budget of 40.0%. [Reason not supplied.
Question to check: did the large trade orders carry a lower margin?]
Freight was $3,900 over budget after the carrier's fuel surcharge
increased from 1 September (Mark).

The output earns its place because nothing in it is invented. Wages, at 2.6% over, fall below the threshold and are left out. The margin question is offered as a question, and the retail shortfall is flagged for the retail manager instead of being explained away.

Where AI management reporting goes wrong

• Invented reasons. The most common and most damaging failure. The placeholder rule is the control.
• Wrong labels on costs. A positive variance on an expense line is unfavourable. Assistants mislabel this often; check every F and U.
• Numbers that do not tie. Rounding, a missed line or a total row counted twice. Tie the draft back to the source report every month.
• GST and period mix-ups. P&L figures usually exclude GST while bank balances include it, and month and year-to-date columns get swapped. State both in the prompt.
• Soft language for bad news. “Slightly softer retail” hides a 12% miss. Ban adjectives and let the numbers speak.
• Automation bias. MBIE’s responsible AI guidance warns that people reviewing AI output can “accept outputs as correct without critical evaluation”. The tie-out step exists because reviewers get tired.

As listed on 19 October 2026, Xero’s AI page describes its JAX assistant and a Claude connector that is read-only, which suits reporting; see our guide to AI and Xero for the options. For the analysis skills underneath, revisit analysing a spreadsheet with AI. Teams producing this every month can compare our management reporting commentary and month-end reporting preparation workflows, and our AI reporting service builds the same controls into a repeatable process.

Self-check

• Every figure in the draft ties to your P&L for the right month.
• You recomputed two variances and they matched.
• Every favourable and unfavourable label is correct, especially on cost lines.
• Every explanation names the person who supplied it.
• No material variance is explained without a note; gaps are marked and chased.
• The commentary contains no vague adjectives standing in for numbers.
• The person who signs the report has read it in full.

Next lesson: Using AI for customer support drafts.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.