To turn a monthly accounts export into a management pack with ChatGPT, agree the reporting basis, map accounts to a P&L, and have it calculate tables before drafting commentary or slides. Reconcile every subtotal to the ledger and approved budget, verify variance explanations with the responsible person, and label unconfirmed causes.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA management pack is a decision document, not a prettier accounts export. Keep the numbers, the explanations and the proposed actions separate until each has been checked. An accurate profit figure does not prove that cash improved, and a plausible explanation for a budget gap is not evidence of its cause.
What should a monthly management pack answer?
A monthly management pack should answer only the questions the meeting needs. For the HR consultancy used as the example here, there are three: did delivery make the expected profit, which gaps need investigating, and which cash or collection issue needs an owner? That scope suggests five pages: a P&L, a budget bridge, a cash and receivables panel, a short commentary and an action list. It does not need every ratio a model can produce.
Agree the reporting period, currency, accounting basis, budget version and close status with whoever is responsible for the accounts. Use a closed ledger, or one clearly marked provisional, with account codes and source references. An export taken before a late adjustment and a budget file saved after a reforecast can both be genuine and still be unsuitable to compare without a version note.
The ledger stays the source of record. ChatGPT is a preparation and drafting tool in this workflow; accounting policy, revenue recognition and adjustments stay with the responsible finance person. OpenAI's own help article for its spreadsheet add-in states that ChatGPT is not a financial, legal or tax advisor and that its outputs should not be treated as advice, which is the right framing for a pack too.
The monthly management reporting tutorial covers the recurring process with any AI tool. This one is the ChatGPT worked example: one export, one budget, and a chain from source records to the meeting pages that someone else can reproduce. Write those inputs on a control sheet before asking for any calculation.
Which files and account mapping does ChatGPT need?
ChatGPT needs three inputs: the current-month accounts export, the approved budget and the previous month's export. Cash and receivables need their own reports, because a P&L alone cannot give closing balances. Remove client names, bank details, personal payroll information and case notes; account codes and anonymised customer IDs are enough for the pack.
ChatGPT reads spreadsheets including XLSX, XLS and CSV, and OpenAI's data analysis guidance asks for descriptive column headers in the first row, one row per record, no unrelated tables on the same sheet, no empty rows or columns splitting the data, and no values trapped in images. That rules out the formatted P&L report most accounting packages print, with its subtotal rows and blank spacer lines. Export a trial balance or account-level detail instead: account code, account name, period, amount.
Plan limits apply. On the pricing page, data analysis is limited on Free and included from Go upwards, and OpenAI's upload help says Free accounts are limited to three file uploads a day while the general allowance is up to 80 files every three hours (OpenAI may lower both at peak times), with spreadsheets capped at roughly 50 MB depending on row size. Three files a month fits comfortably on any paid plan; on Free, upload all three in one go.
Create a mapping table with account code, account name, management category, sign convention and approver. The example consultancy maps subcontractor accounts to Direct delivery costs and office rent to Premises. The category names are your reporting design, not a prescribed chart of accounts. Keep account codes intact so a renamed category never breaks the link back to the ledger.
Sooner or later an expense, say a $900 software subscription, arrives under a new account code the mapping does not know. Put it in an Unmapped queue and stop until finance assigns it. Dropping it into Other would keep total expenses right while distorting the software variance, so a mapping has to pass both a monetary check and a category review.
Test the sign convention on known records. An export may store revenue as credits (negative numbers) while the pack shows revenue as positive. Document the approved conversion and test it on one revenue account and one expense account. Never flip every negative to positive: a reversal or credit note legitimately reduces a category and must keep that meaning.
How do you build the P&L before any commentary?
Build the P&L as a calculation table from the export, showing the source account codes and the arithmetic behind each subtotal, before asking for any prose.
The main example uses invented monthly figures. Revenue is $80,000 and direct delivery costs are $30,000, leaving $50,000 gross profit. Staff, premises, software and other operating costs total $40,000, leaving $10,000 operating profit. The approved budget expected $85,000 of revenue and $15,000 of operating profit.
| Management line | Actual | Budget | Previous month |
|---|---|---|---|
| Revenue | $80,000 | $85,000 | $76,000 |
| Direct delivery costs | $30,000 | $32,000 | $29,000 |
| Gross profit | $50,000 | $53,000 | $47,000 |
| Staff | $28,000 | $27,000 | $27,000 |
| Premises | $4,000 | $4,000 | $4,000 |
| Software | $3,000 | $2,800 | $2,800 |
| Other operating costs | $5,000 | $4,200 | $4,200 |
| Total operating costs | $40,000 | $38,000 | $38,000 |
| Operating profit | $10,000 | $15,000 | $9,000 |
Ask for a calculation table that shows the source account codes and the arithmetic behind each subtotal, with actual, budget and previous month all using the same mapping. Check by hand that $28,000 plus $4,000 plus $3,000 plus $5,000 makes $40,000, and that gross profit less those costs leaves $10,000. If the export contains its own total row, adding it as well would double operating costs and break that check.
Use the attached trial balance export, approved mapping and budget only.
First list each file, its period, row count, sign convention and any
unmapped account codes. Then build the monthly P&L with source account
codes for every line, and show the calculations for gross profit,
operating costs and operating profit. Show the code you used.
Do not write commentary until I approve the table.
The last instruction matters. OpenAI explains that for some analysis ChatGPT writes and runs Python code, and advises reviewing that code, its outputs and its assumptions before relying on the result. You do not need to read Python fluently to spot a filter that drops account 7400 or a sum over the wrong column; asking for the code puts those choices on the screen.
A good first reply shows the nine lines above, each traceable to account codes. Reject an answer that swaps net cash movement for operating profit, leaves out the $5,000 Other category or labels the budget as last month's actual. Those are reporting-basis errors, however neat the formatting looks.
If the reply covers fewer rows than the file holds, do not assume the rest look the same. OpenAI's FAQ notes that a file can upload successfully yet still be too large, complex or poorly structured for complete analysis, and suggests asking about specific sheets or splitting the file. Compare the row count it reports with the count in the export every month.
If the trial balance already lives in Excel, the same table can be built inside the workbook with the ChatGPT for Excel add-in, whose default Skills include financial modelling and corporate finance formatting. The checks do not change: account codes on every line, totals reconciled to the ledger.
If the spreadsheet side is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.
How do you explain a profit gap against budget?
Explain the profit gap with a signed bridge that labels each movement favourable or adverse, so the bridge reads easily. Revenue is $5,000 below budget, which reduces profit by $5,000. Direct costs are $2,000 below budget, which adds $2,000. The gross-profit gap is therefore $3,000 adverse. Operating costs are $2,000 over budget, taking the operating-profit gap to $5,000 adverse.
Split the extra operating costs into $1,000 staff, $200 software and $800 other; premises has no variance. The bridge runs: budget profit $15,000, less $5,000 revenue, plus $2,000 direct-cost saving, less $1,000 staff, less $200 software, less $800 other, giving actual profit of $10,000. Have someone reproduce that sum without looking at the AI's version.
Do not call lower direct costs an efficiency gain without more evidence. The business may simply have delivered less work because revenue was lower. Actual gross margin is $50,000 divided by $80,000, or 62.5%; the budget margin is $53,000 divided by $85,000, about 62.4%. A margin that barely moved calls for a different conversation from a confident story about cost control.
State the denominator for every percentage. Operating profit is one-third below the $15,000 budget, about 33.3% adverse, and about 11.1% above the previous month's $9,000. Both are true, measured against different things. A line such as “profit improved by 33%” would reverse the budget result and hide its basis.
When a category has a zero budget and $250 of actual spend, mark the percentage variance as not meaningful instead of inventing one, show the $250 adverse amount and ask for an explanation. When the comparison value is negative, use the amount and a written explanation rather than a percentage readers will misread. For the wording of variance notes in general, the budget versus actual variance tutorial goes further.
For the bridge chart itself, OpenAI's help says interactive charts cover bar, line, pie and scatter types and other chart types may come back as static images, and the pricing page lists interactive tables and charts on Plus and Pro but not on Free or Go. A waterfall-style bridge is therefore likely to arrive as a picture, so check its bars against the table rather than trusting the image.
How do you stop ChatGPT inventing causes in the commentary?
Stop invented causes by asking, once the tables are approved, for three kinds of statement and keeping them apart: a numerical observation, a confirmed explanation and an open question. “Revenue was $5,000 below budget” is an observation from the table. “A delivery moved into next month” is a cause that needs evidence from the person responsible or the source record. “Check whether timing explains the shortfall” is an investigation task.
Draft five variance notes from the approved table.
For each give: amount, comparison basis, source reference, confirmed
cause or unconfirmed hypothesis, owner, evidence needed, decision needed.
Do not infer customer loss, hiring, price changes or late payments
from totals. Keep proposed actions separate from confirmed facts.
A typical first draft says: “Software costs rose because the team added licences.” The export only proves a $200 adverse variance. The corrected note reads: “Software costs were $3,000 against a $2,800 budget. Finance will check the supplier invoice and renewal schedule before attributing the difference.” When the cause is confirmed, add the evidence reference and update the wording.
Staff costs need the same care. A $1,000 adverse variance could be timing, extra hours, a posting correction or something else, and the model cannot know which. Ask it to list the question for finance and the evidence needed, and keep personal payroll detail out of the pack unless every recipient is entitled to see it.
Set a review threshold and label it as internal policy. This consultancy reviews every line more than $500 away from budget in either direction, plus any smaller item that changes a decision. That pulls in revenue, direct costs, staff and other operating costs this month, while software gets a one-line note. Thresholds save meeting time only if anything important can still be escalated.
If you want help fitting a reporting routine like this into how your team closes the month, I offer one-to-one AI consultations for small businesses.
Why does cash need its own reconciliation?
Cash needs its own reconciliation and its own language, because a fall in cash can sit beside an operating profit. In the cash panel, the verified opening bank balance is $18,000. Receipts are $74,000 and payments $78,000, so the closing balance is $14,000. That $4,000 fall in cash and the $10,000 operating profit are different measures; show both, name the source report for each, and do not let the model blend them into one success story.
A simplified receivables bridge starts at $22,000, adds $80,000 of customer invoices and subtracts $74,000 received, ending at $28,000. In this example there are no other movements. In a real business, finance adds credit notes, write-offs and other adjustments, and the simplified equation never replaces the actual receivables reconciliation.
Together the two panels raise the useful owner question: which unpaid invoices need chasing, and when is the money expected? They do not prove any particular customer is in difficulty. Keep customer-level ageing in a restricted supporting schedule, and put only the level of detail the meeting needs on the slides.
For forward-looking cash decisions, use the 13-week cash forecast workflow. A closing bank balance is a point-in-time figure; a forecast needs assumptions about future receipts and payments, and those belong in a clearly labelled forecast panel, never mixed into the historical actuals.
Catch export errors that balanced subtotals conceal
A duplicated export can double revenue and expenses while leaving the margin looking normal, so compare transaction IDs and row counts with the source report, not just ratios. One duplicated $1,200 subcontractor invoice lifts direct costs from $30,000 to $31,200 and cuts operating profit to $8,800. A polished note about weaker delivery performance would then rest on a data error.
Mixing detail lines with report total rows is the classic trap. If a sheet holds revenue account lines adding to $80,000 plus a Revenue total row of $80,000, summing the column gives $160,000. Decide whether the export is transaction detail, account balances or a formatted report before aggregating, ask ChatGPT to list every row it excluded and why, and look at those rows yourself.
Cut-off problems keep totals plausible while moving performance between months. A delivery finished after month end might carry an earlier draft invoice date. The accounting owner decides the treatment on the approved basis; ChatGPT should never move it to whichever month flatters the comparison. Record any adjustment and whether the ledger itself has been updated.
Rounding causes small display differences. Calculate at full precision and round only for the pack. If three displayed values appear to add to $9,999 while the true total rounds to $10,000, add a rounding note or show one more digit; never change a source amount by a dollar to make a slide tidy. When staff costs need deeper checking, the payroll-to-ledger reconciliation tutorial covers that as supporting work, so the pack itself stays short.
Turn approved numbers into a five-slide meeting pack
OpenAI offers three routes to slides, and the controls are the same for each. ChatGPT Work can create decks and reports from your files and context; OpenAI lists it on all plans in the desktop app (limited on Free and Go), and on Plus, Pro, Business, Enterprise and Edu on the web and mobile. The ChatGPT for PowerPoint add-in, installed from Home > Add-ins in PowerPoint and available on every plan from Free upwards, works inside an existing deck and keeps the slides editable. ChatGPT Slides, on Pro, Business and Enterprise, builds decks you can export to PowerPoint or Google Slides.
For the example, slide one gives revenue of $80,000 and operating profit of $10,000 with the period and a provisional or final label. Slide two is the P&L. Slide three explains the $5,000 gap to budget. Slide four shows cash of $14,000 and receivables of $28,000 with their bridges. Slide five lists decisions and owners, including the unresolved software variance.
Give every slide a source note pointing to the control workbook, and tell the model to use the approved figures exactly rather than recalculate rounded values for decorative charts. If a chart shows $10,000 of profit and its caption says $15,000, stop and resolve it. The budget should be labelled as budget wherever it sits beside actuals.
Then review the exported slides themselves: clipped labels, minus signs, units, comparison titles, and speaker notes that turn an unconfirmed hypothesis into a fact. Check each table against the approved workbook. A slide-generation step can introduce its own transcription errors, so reading the chat reply is not enough.
Approve the pack with a compact control record
Keep a control record with the pack. Here is the consultancy's, filled in:
| Control | This month |
|---|---|
| Source files | Trial balance export, budget v3, previous-month export |
| Ledger status | Closed; no adjustments after export |
| Mapping version | v7, one new code mapped to Software |
| Rows read | Matches the export's row count for each file |
| Revenue and costs agree to ledger | Yes: $80,000 and $70,000 |
| Bridge explains full variance | Yes: $15,000 budget to $10,000 actual |
| Cash and receivables | $14,000 bank, $28,000 debtors, each with its own bridge |
| Open questions | Software variance cause; staff variance cause |
| Prepared by / approved by | Office manager / finance lead |
Where you can, separate the person who prepared the draft from the person who approves the numbers. Where one owner does both, repeat the critical sums outside the AI's answer and note that you did. Before sign-off, run through the list below.
- Every account is mapped or explicitly unresolved.
- Revenue and expenses agree to the approved ledger report.
- Budget and comparison figures use the same definitions.
- The profit bridge explains the whole variance.
- Cash and receivables have their own checked controls.
- Commentary separates confirmed causes from open questions.
- Slides match the approved tables and carry the right revision.
After the meeting, put only agreed decisions on the action register. “Review software variance” becomes an owner, an evidence request and a due date; “reduce costs” is too vague to complete. File the approved pack with its control record, and when a later correction arrives, issue a labelled revision that explains the changed figure rather than quietly replacing the file the meeting used.
Further reads
- How to Speed Up Month-End Close in a Small Business With AI — Close the month sooner so the pack starts from final numbers.
- ChatGPT for Excel: Clean and Summarise a Monthly Sales Export — Clean the sales export behind the revenue line.
- How to Create a Sales Dashboard in Excel Using ChatGPT — Present verified commercial measures on one page.
- What Is Cash Flow Forecasting? A Plain-English Guide With AI Examples — Understand the cash forecast that sits beside the pack.
- Confidential Information Memorandum AI: A Business Broker's Workflow — A broker's workflow for drafting a CIM with AI: outline, financials into narrative, a line-by-line fact-check and confidentiality rules, plus Flippa's builder.
Browse all AI-for-business tutorials
Sources: OpenAI help articles Data analysis with ChatGPT, Uploading files and audio to ChatGPT, and ChatGPT for Excel and Google Sheets; ChatGPT Work, ChatGPT for PowerPoint and ChatGPT Slides pages; ChatGPT pricing page. Checked 18 November 2026.
