Board report from Excel with Claude, ChatGPT or Copilot: KPIs and checks

A quarterly board report built from an Excel workbook can be drafted with Claude for Excel, ChatGPT data analysis or Copilot in Excel, but the numbers in the draft are only as reliable as the checks the finance team runs on them. All three tools can read a profit and loss table, compute margins and variances, and turn them into a written summary. Two of the vendors say so in their documentation. Anthropic states that Claude for Excel is not recommended for final client deliverables without human review, and Microsoft advises against relying on Copilot for decisions in financial matters.

The practical approach is to treat the AI as a fast first drafter and to give it a workbook in which every board figure can be recomputed. The worked example below uses an illustrative software company with a quarterly P&L, a budget, a prior-year column and a cash balance. It lists the prompts to use with each tool and the correct results for every KPI, so a controller can compare the AI's draft line by line before the pack goes to directors.

The most frequent problems in AI-drafted board summaries are arithmetic presented in the wrong unit, cost variances reported with the wrong favorable or unfavorable sign, and explanations of variances that nobody in management actually gave. Each of these has a specific check, described in the review section.

··········

WHAT A QUARTERLY BOARD REPORT NEEDS FROM THE FINANCE TEAM.

Directors usually receive a short financial summary on the first pages of the board pack, followed by detailed schedules. The summary answers a small number of questions: how revenue and profit compared with budget and with the same quarter of the prior year, how much cash the company holds and how fast it is using it, and what management expects next. The detailed schedules support those statements, and the summary must agree with them to the dollar.

For an AI tool, the task has two separate parts. The first is calculation: gross profit, gross margin, operating expenses, EBITDA, EBITDA margin, budget variances, year-over-year growth, cash burn, runway and days sales outstanding. The second is narrative: a few paragraphs that describe the results and the drivers behind them. The calculation part can be verified mechanically. The narrative part can only be verified against what management has said, which is why the source workbook should contain a column of management notes and the prompt should forbid the AI from adding causes of its own.

Teams that already prepare a monthly variance file can reuse it. The approach in the budget vs actual variance analysis with Copilot in Excel produces exactly the variance columns a board summary needs, and the quarterly view is the same calculation over three months.

··········

CLAUDE, CHATGPT AND COPILOT IN EXCEL: DOCUMENTED REQUIREMENTS FOR BOARD REPORTING.

The three tools work in different places. Claude for Excel and Copilot in Excel run inside the workbook. ChatGPT data analysis works on an uploaded file in a separate Python environment. The table summarizes what each vendor documents, as verified on the official documentation pages on November 6, 2026.

........

Item	Claude for Excel	ChatGPT data analysis	Copilot in Excel
Plans	Pro, Max, Team, Enterprise	Not stated on the help page	Microsoft 365 Premium, commercial Copilot, eligible business plans, or Personal/Family with AI credits
Where it runs	Add-in in Excel on the web, Windows, Mac	Uploaded file, Python with pandas	Inside Excel, edit, plan or chat mode
File limits	Claude account usage limits apply	10 files per conversation; 512 MB per file; spreadsheets about 50 MB	Standard .xlsx; Strict Open XML not supported
Useful for board packs	Cell-level citations; can write to open PowerPoint or Word files	Charts and tables computed in Python	Summaries, trends, outliers, PivotTables
Documented limits	Data tables, macros and VBA not supported	No outbound network access	Editing needs automatic calculation

........

Claude for Excel answers questions about a workbook with cell-level citations that link to the referenced cell, which is the most direct way to trace a board figure back to its source. Anthropic also documents a cross-app mode in which Claude reads data from an open Excel workbook and writes it into an open PowerPoint presentation or Word document, for example pulling numbers from an Excel model into a slide or drafting a Word memo from workbook data. Cross-app work requires a paid Claude plan, the Excel, PowerPoint, Word and Outlook add-ins installed, and the "Let Claude work across files" setting, which is on by default for Pro and Max and off by default for Team and Enterprise. Claude can only work with files that are already open; it cannot open, create or switch files itself.

ChatGPT data analysis accepts Excel, CSV, PDF and JSON files and runs Python code with pandas for calculations and Matplotlib for charts. OpenAI documents that spreadsheets work best with descriptive plain-language headers, one record per row, and no multiple tables or empty rows on the same sheet. A board workbook usually breaks those rules on its presentation tabs, so the file uploaded to ChatGPT should be a clean data tab rather than the formatted pack.

Copilot in Excel opens in edit mode by default, and also offers a plan mode, which proposes the steps before changing anything, and a chat-only mode, which analyzes data without changing the workbook. For board work, chat-only mode is the safer choice on the final file, because Microsoft notes that once Copilot changes are saved, anyone with access to the file can see them.

··········

BOARD REPORT WORKED EXAMPLE: QUARTERLY P&L, BUDGET AND PRIOR YEAR.

The dataset below is illustrative. It represents a software company reporting its third quarter of 2026, with figures in USD thousands. The expense lines are assumed to exclude depreciation and amortization, so EBITDA is revenue minus cost of revenue minus operating expenses. The management notes column contains the only explanations the AI is allowed to use.

........

Line item (USD thousands)	Q3 2026 actual	Q3 2026 budget	Q3 2025 actual	Management note
Revenue	4,820	5,100	4,150	One enterprise contract signed in October instead of September
Cost of revenue	1,880	1,940	1,700	Lower hosting volume
Sales and marketing	1,240	1,180	1,050	Trade show moved from Q4 to Q3
Research and development	760	800	690	Two engineering hires started later than planned
General and administrative	520	500	470	Higher audit fee

........

A second small table on the same sheet holds the balance sheet inputs: cash at the start of the quarter of 6,950, cash at the end of the quarter of 6,300, and accounts receivable at quarter end of 2,140. The quarter has 92 days, which is the day count used for days sales outstanding.

The prompt for Claude for Excel asks for calculations in new cells and citations for every figure, so each number in the draft can be clicked back to its source.

On the sheet "Q3 data", add a KPI block below the tables with formulas (not typed values) for:
gross profit, gross margin %, total operating expenses, EBITDA, EBITDA margin %,
variance vs budget in USD and %, change vs Q3 2025 in %, and margin changes in percentage points.
Add cash change for the quarter, average monthly net cash burn, runway in months at that burn,
and DSO using 92 days. Then write a 150-word board summary in chat.
Use only the explanations in the "Management note" column. If a variance has no note, write
"cause to be confirmed by management". Cite the cell for every number you use.

The same request works in ChatGPT after uploading the data tab as an .xlsx or .csv file. The prompt asks for the Python code to be shown, so the calculation can be read and checked.

The attached file contains a quarterly P&L in USD thousands with actual, budget and prior-year columns,
plus cash and receivables. Compute the KPIs listed below with pandas, show the code, and return
a table with actual, budget, variance (USD and %), prior year and change vs prior year.
Report margin differences in percentage points, not percent. Treat a cost above budget as unfavorable.
Then build an EBITDA bridge from budget to actual by line item and check that it sums to actual EBITDA.
Draft a board summary using only the management notes provided.

In Copilot in Excel, chat-only mode keeps the review copy unchanged while the summary is drafted. Microsoft's guidance is that naming the columns to analyze produces more accurate results, so the prompt names them.

Using the columns "Q3 2026 actual", "Q3 2026 budget" and "Q3 2025 actual", summarize revenue,
gross margin, operating expenses and EBITDA against budget and prior year. Express margin changes
in percentage points. List any line where actual differs from budget by more than 3%.

When the summary needs to go into slides, Claude's cross-app mode can take the reviewed KPI block from the open workbook and place it in the open board presentation. The transfer should happen only after the figures have been checked, because the slide then carries the same numbers into a document directors will keep.

··········

EXPECTED BOARD KPIS: THE CORRECT RESULTS TO CHECK THE AI AGAINST.

The results below were computed in Python from the illustrative dataset. Percentages are rounded to one decimal place, and margin differences are in percentage points (pp).

........

KPI (USD thousands)	Actual	Budget	Variance vs budget	Q3 2025	Change vs Q3 2025
Revenue	4,820	5,100	-280 (-5.5%)	4,150	+16.1%
Gross profit	2,940	3,160	-220 (-7.0%)	2,450	+20.0%
Gross margin	61.0%	62.0%	-1.0 pp	59.0%	+2.0 pp
Operating expenses	2,520	2,480	+40 (+1.6%), unfavorable	2,210	+14.0%
EBITDA	420	680	-260 (-38.2%)	240	+75.0%
EBITDA margin	8.7%	13.3%	-4.6 pp	5.8%	+2.9 pp

........

The EBITDA bridge from budget to actual must add up exactly. A favorable variance increases EBITDA and an unfavorable one reduces it.

........

EBITDA bridge step	USD thousands	Effect
Budget EBITDA	680	Starting point
Revenue shortfall	-280	Unfavorable
Cost of revenue below budget	+60	Favorable
Sales and marketing above budget	-60	Unfavorable
Research and development below budget	+40	Favorable
General and administrative above budget	-20	Unfavorable
Actual EBITDA	420	Ending point

........

The cash figures complete the summary. Cash fell by 650 during the quarter, from 6,950 to 6,300, which is an average monthly net cash burn of about 217. At that rate the 6,300 balance covers about 29.1 months. Days sales outstanding are 40.9 days, calculated as receivables of 2,140 divided by quarterly revenue of 4,820, multiplied by 92 days.

A correct draft would read along these lines: revenue of USD 4.82 million was 5.5% below budget and 16.1% above the third quarter of 2025, with the shortfall attributed by management to one enterprise contract signed in October instead of September. Gross margin was 61.0%, 1.0 percentage point below budget and 2.0 points above the prior year. EBITDA of USD 0.42 million was USD 0.26 million below budget, mainly because of the revenue shortfall and the trade show moved into the quarter, partly offset by lower hosting costs and later engineering hires. Cash ended the quarter at USD 6.3 million.

··········

REVIEWING AN AI-DRAFTED BOARD SUMMARY: ERRORS AND CONTROLS.

Margin changes are the most common unit error. Gross margin moved from a budgeted 62.0% to an actual 61.0%, which is a decline of 1.0 percentage point. An AI that divides the two margins reports a decline of about 1.6%, a figure that is arithmetically correct as a relative change and misleading in a board summary. The prompt should name the unit, and the reviewer should check every margin statement for "pp" or "percentage points".

Cost variance signs are the second error. Operating expenses of 2,520 against a budget of 2,480 give a positive difference of 40, which is unfavorable. Tools that apply one sign convention to every row describe the overspend as an improvement. The EBITDA bridge is the control: if the steps do not sum from 680 to 420, at least one sign is wrong.

The third error is invented commentary. Asked for drivers, a model will often supply plausible ones, such as pricing pressure or seasonality, that appear nowhere in the source. Each sentence of the narrative should be matched to a management note or marked as a cause to be confirmed. Cash needs the same care: the example shows positive EBITDA of 420 and a cash decrease of 650, and the difference comes from items the dataset does not contain, such as working capital movements or capital expenditure. A draft that explains the gap without data in the workbook has made the explanation up.

Runway and DSO definitions should match the company's own policy. Runway at the average burn of the last quarter is one convention among several, and DSO can be computed on quarter-end or average receivables. The board pack should state the definition used, and the AI should be told which one applies. For valuation-type sensitivities, such as testing how EBITDA or value changes when assumptions move, the stress-testing steps in the DCF model in Excel with Claude or ChatGPT apply directly; Anthropic documents that Claude for Excel does not support Excel data tables, so those grids are built with formulas instead.

A short control list for the reviewer covers the essentials:

........

Control	How to check	Owner
Totals tie to the ledger	Revenue and expense lines agree to the trial balance	Controller
KPIs recomputed	Formulas in the KPI block, or a Python recalculation	Financial analyst
Units and signs	Margins in pp; cost overruns unfavorable; bridge sums	Reviewer
Commentary sourced	Every driver matches a management note	CFO
Final figures locked	Slides and memo use the reviewed values only	Controller

........

The control list follows the same logic as the review steps in month-end close with Copilot in Excel, Claude and ChatGPT: the AI prepares, and a named person approves. For teams that present the same figures in Power BI, the variance measures described in the Power BI budget vs actual dashboard with Copilot and DAX give an independent recalculation of the same KPIs.

··········

DATA PRIVACY FOR BOARD MATERIAL IN CLAUDE, CHATGPT AND COPILOT.

Board packs contain non-public financial results, so the data-handling terms of each tool should be checked before the first real workbook is used. Anthropic documents that when a user signs in with a Claude account, the add-in sends conversations to Anthropic, including the content Claude reads from the workbook and any attached files. Inputs and outputs are deleted from the backend within 30 days of receipt or generation, with the exceptions listed in Anthropic's privacy documentation. Claude for Excel does not inherit custom data retention settings an organization may have configured, and its activity is not included in Enterprise audit logs, although Enterprise organizations with the Compliance API enabled have add-in sessions covered by it. Chat history sits locally in the browser and clearing it does not delete what Anthropic keeps.

Anthropic also warns that files from external sources can contain hidden instructions that manipulate the add-in, and recommends using Claude for Excel only with trusted spreadsheets. A board workbook assembled from subsidiary or vendor files should be rebuilt as a clean internal data tab before the AI reads it.

OpenAI documents that each ChatGPT conversation runs in its own isolated execution environment, destroyed within 13 hours after a conversation goes inactive. Files in chats are kept for the retention period of the parent chat, and files of deleted chats are removed from OpenAI's systems within 30 days, with exceptions for de-identified data and security or legal reasons. Files processed through data analysis are kept for a period that varies by plan, and the help page does not give the durations.

Microsoft notes that saved Copilot changes are visible to anyone with access to the file, and that users can view past Copilot conversations but can only delete their full Copilot activity history, not individual conversations. For a board file with restricted access, working in chat-only mode on a review copy avoids leaving AI edits in the version directors receive.

··········

AI BOARD REPORTING FROM EXCEL: WHAT THE TOOLS DO AND WHAT STAYS WITH FINANCE.

All three tools can compute the standard board KPIs from a clean quarterly P&L and draft a readable summary in minutes. Claude for Excel adds cell-level citations and a documented path from the workbook to an open PowerPoint or Word file. ChatGPT shows its Python code and produces charts outside the workbook. Copilot in Excel works on the file itself, with a chat-only mode that leaves the workbook unchanged.

In the illustrative quarter, the correct draft reports revenue 5.5% below budget and 16.1% above the prior year, gross margin down 1.0 percentage point against budget, EBITDA of 420 against a budget of 680 with a bridge that sums exactly, and a cash balance of 6,300 with about 29 months of runway at the third-quarter burn rate. Those are the figures to compare with any AI output before it reaches the board.

The parts that stay with the finance team are the definitions, the explanations and the approval. Which runway and DSO conventions apply, why revenue missed budget, and whether the summary is fit to send are decisions the vendors' own documentation leaves to human review.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org