Quick answer
To write a board report using AI safely, put every figure into one structured table first, then instruct the model to describe only those values and never calculate or explain a number that is missing. Draft each board section separately, and reconcile every figure before circulation.
Writing a board report using AI is the fastest way to turn a fifteen-tab workbook into two readable pages — and the fastest way to put a figure in front of your directors that nothing in your accounts actually supports. Both outcomes come out of the same tool on the same afternoon. The difference sits entirely in how you hand over the data.
Most people paste a messy block of numbers and departmental notes into a chat window, ask for an executive summary, and get back something that reads beautifully and quotes a variance nobody can trace. That is an input failure rather than a model failure, and it takes roughly ten minutes of setup to close permanently. What follows is the Locked Table Method, the approach we teach at PromptPeakAI for any document where a wrong number has consequences:
- Lock — every figure goes into a structured table before AI sees any of it.
- Frame — the model is told who reads this and what decision it supports.
- Draft — each board section is written on its own, never all at once.
- Reconcile — every number in the draft is matched back to the locked table.
Four stages, six copy-paste prompts, and one rule doing most of the work.
Will this work for you?
Which Account May See Board Data
Before you write a single prompt, settle the account question. Anyone writing a board report using AI should be doing it inside a business or enterprise workspace their employer controls, never a personal account. The boundary is the account type, not the price you pay for it.
This is the single most common mistake in guides on this topic, and it is worth being blunt about. Upgrading a personal subscription changes your usage limits. It does not change how your inputs are handled by default.
The line that matters runs between account types, not between free and paid plans.
What the vendors actually say
Microsoft states plainly in its enterprise data protection documentation that prompts, responses and data reached through Microsoft Graph are not used to train foundation models, and that your existing permissions, sensitivity labels and retention policies carry over into Copilot. OpenAI draws the same line: its Business workspace documentation confirms workspace data is excluded from training by default and encrypted in transit and at rest.
Personal plans are governed by consumer terms instead, where training is a setting an individual can toggle. That is a fine arrangement for your own drafts. It is the wrong arrangement for pre-release earnings.
Public filings, your own prose, anonymised structures with stand-in figures.
Named customers, named employees, supplier terms, anything identifying a person.
Pre-release results, board minutes, valuations, legal matters, personal data.
What to redact anyway
A business workspace protects your data from the vendor. It does not protect it from colleagues, screenshots or a shared link sent to the wrong person. Strip customer names, individual salaries and anything under legal privilege even inside an approved workspace — the report reads the same without them, and your risk profile drops sharply.
Two questions settle this in under a minute, and they are worth asking before your first quarterly run rather than after it:
- Whose account is this? If you signed up with a personal email address, it is a consumer account regardless of what you pay.
- Who set the data policy? On a business workspace an admin sets it centrally. On a personal plan, one distracted click reverses it.
If you want the longer version for a specific tool, we have written it up for both major ecosystems: ChatGPT’s workplace privacy picture and what Copilot can see inside your tenant.
The Locked Table Method Explained
The Locked Table Method is a drafting discipline: every figure that will appear in the board paper is placed in a structured table first, and the AI is instructed to describe only the values in that table — never to calculate, round, restate or explain a number that is not there.
This matters more for a board report using AI than for almost any other document. Language models predict text. They do not do arithmetic reliably on prose, and when a number is missing they will produce a plausible one because plausible text is exactly what they are built for. Give them a locked table and the arithmetic is already done; all that remains is the writing, which is the part they are genuinely good at.
Four stages, each one narrowing what the model is allowed to invent.
The difference this makes is not subtle. Here is the same quarter, handed over two different ways.
✕ Prose in, invention out
"Marketing came in over budget again this quarter, roughly 12 or 13 percent I think, mostly the campaign spend. Write the commentary." Output: "Marketing overspent by 12.8%, driven primarily by an unplanned brand awareness campaign that leveraged higher-cost digital channels in a competitive market."
✓ Locked table in, description out
Table row supplied: Marketing | Budget 400,000 | Actual 451,000 | Variance +12.75% | Note from department: "Q2 campaign brought forward from Q3" Output: "Marketing was 12.75% above budget. The department attributes this to the Q2 campaign being brought forward from Q3."
Notice what changed. The second output is duller, shorter, and every word of it survives a director asking where the figure came from. Dull and defensible is the correct target for a board paper.
Build Your Board Report Using AI
Four stages, in order. The first run takes around an hour because you are building templates you will reuse; every quarter after that lands closer to twenty-five minutes.
Lock every figure into one table
Open a blank document and build a single table containing every number that will appear in the paper. Do the arithmetic in your spreadsheet, not in the chat window. Each row carries the metric, the budget or prior-period figure, the actual, the variance already calculated, and a short note from whoever owns that line.
That last column is the one people skip and the one that saves you. Without it the model has no cause to attach to a variance, so it will supply a cause of its own. If you are pulling the raw numbers out of a workbook, our walkthrough on getting analysis out of a spreadsheet without Excel skills covers the extraction side.
I am preparing a board paper. Below is raw data pasted from a spreadsheet. Convert it into a single markdown table with these columns: Metric, Prior/Budget, Actual, Variance %, Department Note. Rules: - Do not calculate, recalculate, round or adjust any value. Copy them exactly as given. - If a variance percentage is missing, write MISSING rather than working it out. - If a department note is missing, write NOTE REQUIRED. - Add no commentary, no interpretation, no summary. [paste your raw data here]
You should now have one table where every cell is either a real value you supplied or a visible MISSING flag — and no invented figures anywhere.
A markdown financial table pasted into an AI chat box directly above the locked-table constraint prompt, before pressing send.
Crop to the composer area only. Hide the sidebar, the account name, and the workspace switcher. Replace all values with round dummy figures.
File: board-report-locked-table-prompt.png · Alt: "Locked financial table and constraint prompt pasted into an AI chat box for a board report"Frame the reader and the decision
A board paper fails more often on tone than on content. Directors read a large stack of pre-reading before every meeting, and a summary that spends three sentences setting the scene before reaching the request has already lost. Write the frame once, save it, and paste it at the top of every drafting prompt.
CONTEXT FOR EVERY SECTION BELOW: You are drafting for a board of directors at [company], meeting on [date]. Readers: [e.g. six non-executive directors, two of whom are not finance specialists]. The decision this paper supports: [e.g. approve the revised capital plan]. Writing rules: - State the point in the first sentence. No scene-setting opener. - Plain English. No marketing language, no adjectives of praise. - Use only figures present in the table I supply. Never introduce a new one. - Where a cause is not stated in the notes, write [cause requires department input].
Draft one section at a time
Four sections, four separate prompts. Drafting the whole paper in one request is what produces the mushy middle where figures drift between sections — and it makes errors far harder to spot, because you are reviewing everything at once.
Each section has a different relationship to the locked table. Two consume figures, two do not.
Start with the financial commentary rather than the summary. The summary is a distillation of everything else, so writing it first means distilling material that does not exist yet — which is precisely when a model starts filling gaps. Our guide to writing a financial narrative for a leadership audience goes deeper on the variance-explanation craft.
[paste your frame block] Using ONLY the table below, write the Financial Commentary section. Structure: - One opening sentence stating overall position against budget. - Then one short paragraph per variance above 5% in either direction. - Each paragraph: the metric, the variance as stated, the department's own explanation. Do not aggregate, average or recalculate anything. Do not explain a variance the notes do not explain. Maximum 300 words. [paste your locked table]
Risk and outlook come next. Risk is the section where invention is most tempting and most damaging, because a rating you cannot defend is worse than no rating at all. If you are building the underlying register from scratch, our 20-minute risk assessment workflow produces exactly the input this prompt expects, and traffic-light status reporting covers the red-amber-green layer directors scan first.
[paste your frame block] Write two sections from the material below. RISK REGISTER — a table with columns: Risk, Likelihood, Impact, Owner, Mitigation. Use only the risks, ratings and owners I have supplied. Do not add risks you consider likely. Do not upgrade or downgrade any rating. OUTLOOK — maximum 150 words on direction of travel and what the board is being asked to decide. Include no forecast figures. Where a forward number would normally appear, write [forecast to be confirmed]. [paste your risk list and your locked table]
The executive summary goes last, built from the sections you have already approved. Directors read this one and skim the rest, so it carries the request in its first line. The five-bullet executive summary method is the pattern this prompt applies.
[paste your frame block] Below are the finished Financial Commentary, Risk Register and Outlook sections. Write the Executive Summary as five bullets, in this order: 1. The decision or approval being requested, stated in one line. 2. Overall position against plan. 3. The single most material variance. 4. The highest-rated open risk. 5. What happens next and by when. Every figure must already appear in the sections below. Introduce nothing new. Maximum 140 words total. [paste your drafted sections]
You should now have four drafted sections, written in this order: financial, risk, outlook, summary — each one built from material you supplied rather than material the model chose.
Consolidating six regional updates written in six different voices is the other half of many board packs. That is a normalisation job rather than an analysis job, and our guide to synthesising multiple documents into one brief handles it properly. Feed the consolidated result in as narrative context — never as a source of figures.
Reconcile before anyone else reads it
This stage is non-negotiable and takes about five minutes. You are not proofreading the prose; you are checking that every number in the draft exists in the table you locked. Run the audit prompt, then do a human pass over whatever it flags.
Below are (A) my locked source table and (B) a drafted board paper. List every numeric value that appears in B. For each one, state: - The value as written in B - Whether it appears in A, exactly - If it does not appear in A: MISMATCH Do not rewrite the draft. Do not fix anything. Output the list only. A: [paste locked table] B: [paste drafted sections]
You should now have a line-by-line list where every figure resolves to a row you supplied, and any MISMATCH is either corrected by hand or removed before circulation.
Copilot vs ChatGPT vs Claude Compared
When it comes to a board report using AI, the tool matters far less than the account and the method — but the three main options do have genuinely different strengths, and one of them is usually already sitting on your desk.
| Tool | Strongest for board work | Main constraint |
|---|---|---|
| Microsoft 365 Copilot | Drafting inside Word next to the source workbook | Needs a licence your admin controls |
| ChatGPT Business | Structured analysis of uploaded tables and files | Lives outside your document, so more pasting |
| Claude Team | Reading long board packs and prior-year papers | Also outside Word, no native Office drafting |
| Free tiers of any tool | Learning the method on anonymised practice data | Consumer terms, so never real board figures |
Copilot’s advantage is proximity. The paper is being written in Word, the numbers live in an Excel file in the same tenant, and the draft never leaves either. If your organisation has licences, start there — our Copilot in Word walkthrough covers the panel mechanics.
A Word document containing a board paper section, with the Copilot side panel open and drafting commentary from a linked Excel table.
Hide the user profile name, the tenant domain in the title bar, and any real figures. Use round dummy values throughout.
File: copilot-word-board-report-draft.png · Alt: "Copilot side panel drafting board report commentary inside a Word document"ChatGPT Business earns its place when the input is genuinely a data problem — several files, inconsistent formats, a table that needs restructuring before it can be locked. Claude is the one to reach for when the input is length: last year’s board pack, three committee papers, a long audit report you need summarised without losing the caveats. Turning any of those into readable prose is covered in our guide to writing a data narrative from raw numbers.
Honestly, if you already have one of the three, use it. The gap between these tools on this task is far smaller than the gap between a locked table and a paragraph of pasted notes.
What AI Cannot Do Here
Some parts of a board report using AI are not delegable at all, and pretending otherwise is how people get caught out. The model can produce every sentence in the document and still leave the most important work untouched.
Four things remain entirely yours, and no prompt improves them:
- Deciding what the board needs to know. A model weighs every row equally. Materiality is a judgement about consequences, and it is the core of the job.
- Standing behind the numbers. Your name goes on the paper. Fiduciary responsibility does not transfer to a tool.
- Reading the room. The model does not know which director asked about this last quarter, or which topic is politically live.
- Delivering genuinely bad news. Framing a shortfall honestly without either minimising it or triggering panic is a human skill built on relationships.
There is also a category error worth naming. AI is good at making writing clearer; it is not good at making a weak position look strong, and attempting the latter produces exactly the confident, evasive prose experienced directors are trained to distrust.
✕ Asking AI to carry the judgement
"The Q2 shortfall reflects a strategic recalibration of our go-to-market motion, positioning the business to capture disproportionate share as market conditions normalise."
✓ Human judgement, AI clarity
"Revenue finished 9% below plan. Two of the three causes are within our control and are addressed in the actions below. We do not expect to recover the gap this year."
Most experienced company secretaries will tell you the second version buys more credibility in the room, not less. Boards respond well to a management team that names a problem before they have to.
When the Draft Goes Wrong
Four failures account for nearly everything that goes wrong when drafting a board report using AI. Each has a specific cause and a specific fix.
| What you see | Why it happens | The fix |
|---|---|---|
| A figure you cannot trace | Data arrived as prose, so gaps were filled | Rebuild the locked table and rerun |
| Glossy, empty corporate prose | No reader named, no banned words set | Add the frame block above every prompt |
| The draft stops mid-section | Input exceeded what the session holds | Draft one section per prompt, not the paper |
| A cause nobody in the business stated | The notes column was left blank | Add notes, or accept the bracketed flag |
The bracketed-flag habit deserves a word of its own. Instructing the model to write [cause requires department input] rather than guess feels clumsy in a draft, and it is the single most valuable line in the whole frame block. A visible gap gets chased. An invented explanation gets published.
Run this on every number before the paper leaves your desk.
One more pattern worth watching for: a draft that is technically accurate and quietly reorders your priorities, leading with the good news because good news makes for smoother prose. Read the summary once more asking only whether the first bullet is genuinely the most important thing in the pack.
Making It a Quarterly System
The value here compounds. Your locked table structure, your frame block and your six prompts do not change between quarters — only the values do. Save them somewhere you will find them in three months, and the second run costs a fraction of the first.
- Keep the frame block in the same file as the table template.
- Update the reader list whenever the board composition changes.
- Re-run the reconcile prompt even on a quarter that felt straightforward.
If your organisation has not settled the account question yet, do that before the next cycle rather than during it — it is the one prerequisite for a board report using AI that no prompt can work around. Everything else in this method works identically on any of the three tools. The account is the part you cannot fix at eleven at night the day before the pack goes out.
Lock your figures before the next board cycle
- Confirm you are on a work account, not a personal plan Build one table holding every figure, with the arithmetic already done Add a note column naming who explained each variance Write the frame block once and reuse it on every drafting prompt Draft financial, risk, outlook, then the summary last Reconcile every number against your table before circulating