Build Your Board Deck with AI: The 6-Prompt Playbook

TL;DR Six prompts. Copy them in order. Feed a P&L export, a cash file, a Slack digest, and a plan. You get a board-ready deck draft in about 45 minutes instead of two days. You still edit the numbers, own the narrative, and cut the caveats. Everything below is what has worked on real monthly and quarterly PE-backed board packages.

The first draft of a board deck is where finance teams lose whole days. Not building the model. Not closing the books. Writing the story around numbers you already know cold.

An LLM cuts that cycle down if you use it as a structured drafting partner instead of a magic answer machine. The recipe below is six prompts you run in order. Each one produces a chunk of the deck. You review, edit, and paste. The AI writes prose. You own the numbers.

Why six prompts and not one big one

Everyone tries the one-prompt version first. Paste a P&L, ask for a board deck, get back a bland summary that either invents context or restates the numbers you already gave it. It fails because a board deck has six distinct pieces with different inputs, different audiences, and different tolerances for uncertainty. Break them apart and each prompt gets small enough that the model can actually reason well.

You also get a useful side effect. If prompt 3 (cash) is weak, you rerun it in isolation. You are not regenerating the whole deck to fix one section.

What you need before you start

Actuals-vs-plan CSV (P&L level, 12-24 months back)
Current cash position and covenant test results
A dump of Slack channels or email threads for the top 5 initiatives (last 30 days)
The current forecast file with 2-3 sensitivity scenarios
Prior month’s board deck or exec summary for tone continuity

Prompt 1: Variance narrative from actuals vs plan

This is where most decks fail. The temptation is to write “revenue was up 3% due to strong performance” which tells the board nothing. This prompt forces the model to name the driver.

You are an FP&A analyst writing the variance section of a monthly board deck.

Attached CSV: actuals vs plan for {MONTH} at the P&L line level, with prior 3 months for trend.

Write:
1. Three bullet variance drivers for revenue (unit mix, price, timing, or channel). Name the driver, not the direction.
2. Three bullet drivers for gross margin (input cost, mix, discounting, or one-time).
3. Two bullet drivers for opex (headcount timing, T&E, marketing pull-forward, or vendor renewal).

Rules:
- Do not say "strong performance" or "underperformed." Name the mechanism.
- If the CSV does not contain a driver, write "driver not visible in P&L, needs review with {FUNCTION}."
- Use dollar amounts and basis points, not percentages of percentages.
- Maximum 25 words per bullet.

Prompt 2: KPI dashboard structure

Boards read the KPI page first and the appendix last. This prompt builds the table structure. You fill in the numbers by hand or pipe them in from a second file.

Build a KPI table for a {INDUSTRY} PE-backed company at ~{REVENUE{'}'} revenue.

Columns: Metric | This Month | Plan | Prior Month | LTM | Trend Arrow.

Rows (in this order):
1. Revenue
2. Gross margin %
3. EBITDA (adj)
4. EBITDA margin %
5. Cash on hand
6. Net debt
7. Net debt / EBITDA (LTM)
8. Days sales outstanding
9. Days payable outstanding
10. Rolling 3-month net new customer / logo count
11. One industry-specific KPI relevant to {INDUSTRY} (name it)

Output the table in markdown. Below the table, write 3 bullet points on which KPI moved most in the wrong direction and why the board should care.

Prompt 3: Cash and covenant summary

Sponsors care about two things in cash: when do we run out, and are we still inside the credit agreement. This prompt structures both.

Attached: current 13-week cash forecast (CSV) and covenant test spreadsheet with the maintenance ratios from our credit agreement.

Write a cash and covenant section with:

1. One sentence on current cash position and change vs prior month.
2. Runway table: cash on hand, monthly burn (or generation), months of runway at current trajectory.
3. Covenant status: for each maintenance covenant, current ratio vs required, cushion in dollars, and forecast month when cushion narrows to less than 15%.
4. Two bullet risk flags: things that could compress cushion inside 90 days.

Rules:
- No hedging language. If a covenant will break inside the forecast horizon, say so.
- If data is missing from the file, write "not in file, verify with treasury" and stop that section.

Prompt 4: Top 5 initiatives status from Slack and email

This is the prompt that saves the most time. You are pulling qualitative status out of 300 messages instead of chasing five owners for updates.

Attached: text dump of Slack channels and email threads from the last 30 days for these 5 initiatives:
1. {INITIATIVE 1}
2. {INITIATIVE 2}
3. {INITIATIVE 3}
4. {INITIATIVE 4}
5. {INITIATIVE 5}

For each initiative, output:
- Status: On Track / At Risk / Off Track (pick one, no "yellow-green")
- One sentence: what shipped in the last 30 days
- One sentence: what is blocking or slipping
- One sentence: the next 30-day milestone with a date

Rules:
- If you cannot find evidence of progress, write "no signal in inputs, owner check needed."
- Do not infer intent. Report what people wrote.
- Ignore small talk, calendar invites, and out-of-office replies.

Prompt 5: Forecast update with sensitivity

The forecast prompt is where the model wants to invent numbers. Do not let it. Feed it the file and let it write only the story around scenarios you already built.

Attached: current full-year forecast with three scenarios (base, upside, downside). Also attached: prior forecast from last month for delta.

Write the forecast update section:

1. One paragraph on what changed vs prior forecast. Name the top 3 assumption changes with dollar impact.
2. Base case summary: full-year revenue, EBITDA, EBITDA margin, and cash generation.
3. Sensitivity table: for each scenario, show EBITDA delta vs base, single largest assumption driving the difference, and probability weight if I have provided one.
4. One bullet on the assumption we are least confident in.

Rules:
- Do not create new numbers. If a value is not in the file, write "TBD."
- Do not use the word "conservative." Everyone claims conservative. Instead say what specifically changed.

Prompt 6: Executive summary

Run this one last, after you have edited the other five outputs. Paste the finished sections back in and ask for the summary. Otherwise the summary drifts from what actually made it into the deck.

You have my final variance, KPI, cash, initiatives, and forecast sections attached below.

Write a one-page executive summary for a PE board with these five parts, in this order:
1. Headline number (revenue and EBITDA vs plan) in one sentence.
2. What went right (2 bullets, max 15 words each).
3. What went wrong (2 bullets, max 15 words each, name the driver).
4. Where we need help from the board (1-3 asks, each concrete: an intro, a decision, or a hire approval).
5. One-line forward look for next month.

Rules:
- No caveats, no "we continue to monitor."
- If nothing needs board help, write "no asks this month."
- No adjectives on the numbers ("solid," "healthy," "modest"). Just report them.

The review checklist before you send

Check	Why
Every number on the KPI page ties to the source model	LLMs occasionally round or transpose
Every variance driver is named, not described	“Strong performance” gets you 15 minutes of board questions
Covenant cushion is stated in dollars, not just percent	Board members think in absolute headroom
Every initiative status has a date, not “soon”	The AI will default to vague if you let it
Exec summary asks are concrete and answerable in the room	“Feedback welcome” is not an ask
Cash section still reads correctly after your edits	The section most likely to be wrong is the one most read

Three things AI still will not do well for board decks

1. Reading the room. If a board member has been pushing on churn for two months, the model does not know to lead with retention. You do. Reorder the deck accordingly.

2. Political nuance in the “what went wrong” section. The model will name the driver. It will not know that the driver is a sponsor’s favorite hire whose function is missing plan. That framing is human work.

3. Forecasting confidence. An LLM will happily produce a probability-weighted scenario. Treat that number as a placeholder. Real probability weights come from you, your CRO, and your ops lead in a 30-minute call.

Push back on this.

Every operator’s situation is a little different. If you run this differently, disagree with the methodology, or think we got something wrong, tell us. We publish the best counter-approaches on our Reader Contributions page, credited or anonymous, your call. Email hello@thepragmaticcfo.com.

FAQ

Which model handles a P&L CSV best? As of 2026, Claude and ChatGPT both handle a 10-20MB CSV cleanly for structured reasoning. Copilot inside Excel is faster for cell-level work but weaker for narrative. See our Claude vs ChatGPT vs Copilot comparison for the details.

Should I put customer data in the prompt? Not without a private tenant. Rename customer columns to Customer_A, Customer_B before you paste. See the CFO LLM privacy guide.

How long does this actually take? First time, about 3 hours because you are debugging your prompts against your data. By month three, 45 to 60 minutes for the draft, then 90 minutes of edits.

Can I automate it end-to-end? You can. You should not. The point of writing the deck is that you re-encounter every number before the meeting. Automate the drafting, not the reading.

What about the appendix? Appendix is usually last month’s operating detail with light updates. Prompt the model to diff the prior month’s appendix against your current data and highlight what moved by more than 10%.

Sources

Anthropic, Claude documentation on file inputs and long-context reasoning, docs.anthropic.com
AFP, 2026 Treasury Benchmarking Survey (covenant tracking and board reporting practices), afponline.org
AICPA, CFO Advisory practice notes on board package structure, aicpa-cima.com

Written by The Pragmatic CFO. 15+ years running P&Ls and building AI-native finance workflows across portfolio companies.