Chapter 03 · Section I · 17 min read
Drafting a management report from raw figures
The single workflow that saves a Nepali accountant the most report-writing time — and the verification step that keeps you out of trouble.
The management report is where most accountants in Nepal lose the evening. The trial balance closed clean by six, the variance schedule reconciled by seven, and then the partner asks for “a short note for the board, maybe one page, by tomorrow morning” — and three hours of staring at a blank document begins. This is the single piece of work where current AI saves the most time per week for a working accountant, and it is also the single piece of work where a careless prompt produces a report that looks fluent and is quietly wrong. This section is about how to get the time saving without the embarrassment.
The split of labour
The first thing to internalise — and it is the rule that governs this entire chapter — is that your spreadsheet owns the numbers and the model owns the prose around them. Not the other way around. Not a blend. A clean split.
Your trial balance, your P&L, your variance schedule, your budget — these live in Excel, Tally, Zoho Books, or whatever your firm uses. The arithmetic happens there, by formulas you can re-trace, against source documents you can pull from a file. When the report leaves your desk and a number is questioned in the boardroom, you must be able to point at where that number came from and how it was computed. A language model cannot stand behind a number that way; it can produce one, but it cannot defend one.
What the model is genuinely good at is the second half: taking those defended numbers and turning them into a paragraph a non-finance reader can act on. “Revenue 12.4 / 11.1 / 9.8” is a row in a spreadsheet. “Revenue grew 11.7% over Q1 and 26.5% over the same quarter last year, driven primarily by the post-Dashain order book — though gross margin contracted 180 basis points as input costs outpaced our price revisions” is a sentence the board can use. The accountant’s hour becomes ten minutes.
A concrete prompt pattern
Most management-report prompts fail because they ask the model to do too many things at once and do not tell it what it is forbidden to do. A serviceable pattern, refined over a few hundred close cycles, looks like this:
“Here are the P&L figures for Q2 FY 2081/82 for a Nepali manufacturing firm. Columns are actual, budget, prior-year same quarter, all in NPR thousand. Write a 200-word management commentary covering revenue, gross margin, operating expenses, and one risk to flag. Use the figures provided; do not invent any number. If a figure is not in the data, say so explicitly. Audience is the board; tone is sober, not promotional. Provide the same commentary in Nepali underneath.”
Five things are doing the work in that prompt. 1. The figures are pasted in directly — actual, budget, prior year — so the model has all three reference points without having to guess. 2. The output length and scope are bounded: 200 words, four topics, one risk. A model given no length budget produces a meandering page. 3. The forbidden behaviour is stated explicitly: do not invent. 4. The audience and register are named, so you do not get a marketing pitch where you wanted a sober note. 5. The bilingual request is one line, not a separate translation step.
The non-negotiable verification step
Here is the part nobody talks about in the demo videos. The draft the model gives you back will be fluent. It will be persuasive. And roughly one time in five, on a report with several growth rates and ratios, one of the percentages will be wrong. Not wildly wrong — wrong in the second decimal place, or wrong because the model used “growth over prior year” where you wanted “growth over budget,” or wrong because it inverted a margin calculation. The prose will read confidently around the wrong number.
So: before any AI-drafted report leaves your desk, re-compute every percentage, every growth rate, every ratio in the draft against your spreadsheet. This is not a polite suggestion. It is the verification step that separates the accountants who get value from AI from the ones who get a quiet professional disaster eighteen months in. The re-computation takes five minutes. The disaster is permanent.
A practical rhythm: read the draft once for sense, mark every numerical claim with a pencil, and verify each mark against the source. If you find one error, assume there is another and check twice as carefully. The model is not trying to deceive you. It simply does not know when it is guessing.
Where AI is genuinely additive
Once the verification discipline is in place, the productivity gain is real and worth being honest about.
1. Readable prose around defended numbers. Turning “rev 12.4 / 11.1 / 9.8” rows into English or Nepali a non-finance board member can follow. This is the headline use, and it clears the verification test cleanly because the numbers are yours.
2. Executive summary selection. Hand the model the full management pack and ask: “Which three lines from this report would you put in a one-paragraph executive summary for the chairperson?” The model is good at compression. You still pick the final three — but the candidate list saves the thinking time.
3. De-jargoning for SME owners. A Birgunj garments wholesaler does not need “EBITDA margin contracted on input cost pressure.” She needs “your profit per rupee of sales fell because raw material got more expensive faster than your prices.” The model handles this rewrite well. Read it before sending.
4. Restructuring for a different audience. The same underlying numbers go to the board, to the bank, and to the tax adviser in three different forms. The model is fast at re-framing — board gets strategic, bank gets debt-service-coverage, tax adviser gets a clean reconciliation note.
Bilingual reports in one pass
The bilingual request deserves its own paragraph because it is where many Nepali firms see the largest immediate win. Generating the management commentary in English and Nepali in the same prompt is roughly twice as fast as writing one and translating it. The model handles parallel generation well, and the two versions tend to be more consistent than a human translation because they are produced from the same source data in the same pass.
The catch is the same catch as before, doubled: read both versions before sending. Nepali nuance from the model is uneven — honorifics drift, financial terms sometimes get over-anglicised (growth rate instead of वृद्धि दर), and occasionally a number stated correctly in English gets restated incorrectly in Nepali because the model re-typed it. Two languages, two verifications. It is still faster than the old way by a comfortable margin.
Check your understanding
Quick check
—You ask the model to draft a management commentary from the P&L figures you provided. The draft reads beautifully, but on review you notice it claims revenue grew 14.2% when the spreadsheet says 12.7%. What is the correct response?
Quick check
—Which of the following additions would most strengthen a prompt asking an LLM to draft a board commentary from your P&L figures?
What comes next
Drafting prose around your own numbers is the foundational AI workflow for accountants. The next section takes the same discipline into the harder territory of variance analysis — the “why did this change?” question that fills half of every management meeting, and where structured prompting separates a useful candidate list of explanations from a confident-sounding wrong story.