A five-step system for turning messy notes and a metrics spreadsheet into a board update that actually survives questions, built from watching this go right, and watching it go sideways, in the prep sessions I run with founders and managers.
Board prep was already broken before AI showed up. The average board pack now runs 200+ pages, more than double a decade ago, and directors spend just shy of four hours reading it (Board Intelligence and Cambridge Judge Business School). On top of that, 75% of Fortune 500 CEOs and senior executives have already used generative AI for board-related work in the past six months, mostly with no policy telling them how (CEOWORLD). AI earns its keep on the mechanical parts of this job. It structures your notes into a memo, turns a spreadsheet into a narrative, and plays a skeptical board member so you rehearse the hard questions before the real ones land. What it also does, and this part is documented, not anecdotal, is soften specific, uncomfortable detail into vague, agreeable language during summarization. Researchers call it sycophancy. Avoiding AI isn’t the answer. Building one deliberate step, a de-polish pass against your own raw notes, into every prep cycle before anything reaches your board, is.
A Series B founder I worked with ran his board deck through ChatGPT twice before the meeting. Once to tighten the language. Once more to make the churn section “sound more confident.” By the time it reached the board, the deck read cleanly and said almost nothing. Two directors asked the same follow-up question inside the first ten minutes: what’s actually driving the number? He didn’t have a sharp answer, because the AI pass had smoothed his own churn story into something that sounded fine instead of something that was true.
That’s the trap this guide is really about. You should use AI to prep for a board meeting or an investor update, honestly, I’d say start today if you haven’t already. The trap sits one layer down: using it in a way that makes your materials look more finished while your actual grip on the business quietly gets weaker, right up until someone in the room asks a real question.
Here’s the part almost nobody says out loud: board prep was already broken before AI showed up. Board Intelligence’s research with Cambridge Judge Business School found the average board pack now runs 200+ pages, more than double what it was a decade ago, while the average director spends just 4 hours actually reading it (Board Intelligence & Cambridge Judge Business School, In the Boardroom, Size Matters). Do the math and roughly half of what gets sent never gets read with any real attention. That gap existed long before anyone thought to open ChatGPT.
Meanwhile, AI has already shown up in the boardroom whether anyone planned for it or not. A full 75% of Fortune 500 CEOs and senior executives, three out of every four, have used generative AI for board-related work in the past six months, according to a CEOWORLD survey of boardroom executives (CEOWORLD, 94% of Boards Have No AI Policy, Yet 75% of CEOs Are Already Using It). Fewer than one in ten of their boards have a formal policy governing that use. People are already doing this, mostly quietly, without much of a process behind it. Whether you should use AI here isn’t really up for debate anymore. What matters is whether your process holds up once a director starts asking follow-up questions.
Here’s my honest read after watching this play out with founders and managers in the workshops I run: AI handles the mechanical parts of board prep well, the structure, the first draft, turning a spreadsheet into sentences a person would actually say out loud. Ask it to decide what your board actually needs to hear, though, and it’ll happily help you dodge the hard part instead, if you let it.
Skip the tool-by-tool comparison for a second. ChatGPT, Claude, and Gemini all handle this job at a similar level right now, and which one you use matters less than the process you run it through. Here’s the workflow I walk founders and managers through before a board meeting or an investor update.
The five-step workflow described in this guide, from raw notes to a board-ready update.
The next four sections walk through each step in detail, including the one almost everyone skips: the de-polish pass, which is exactly where that founder’s churn number would have been caught before it cost him credibility in the room.
You almost certainly already have the raw material for your board update sitting around somewhere. It’s just scattered: a Slack thread with your head of sales, a half-finished doc from last quarter, maybe a voice memo you left yourself after a rough week. AI is actually useful here. It’s good at pattern-matching structure onto mess, a task that used to eat an entire evening.
Sequoia Capital’s own guidance to its portfolio founders makes a point worth stealing: board decks don’t have to be decks at all. Companies like Qualtrics and Thumbtack run Amazon-style written memos instead of slides, because a memo forces sharper thinking than a slide ever will (Sequoia Capital, Preparing a Board Deck). There’s a reason for that beyond taste: a sentence has to resolve into an actual subject and a verb, while a bullet point just has to look tidy sitting next to five other bullet points. If your board doesn’t require a slide format, try the memo version first. It’s a lot harder to hide a fuzzy point in a paragraph than behind a bullet.
“Here are my raw notes, Slack messages, and a rough metrics export from this quarter. Draft a board memo with these sections: CEO summary, wins, challenges, where I need the board’s help, and key metrics with one line of context each. Keep my actual numbers and specific examples exactly as I wrote them. Don’t generalize or soften anything I flagged as a problem.”
That last instruction matters more than it looks. Without it, AI tends to average your rough notes into something that reads well and says less. Treat the output as a skeleton, not a finished memo. You still have to read every line and check whether it’s actually true, not just whether it sounds true.
Here’s what that looks like in practice. A founder’s raw note might read: “lost the Meridian account, they went with a competitor who’s $3K/month cheaper, our onboarding took too long and their champion left.” A first AI pass, left unchecked, often turns that into: “we experienced some customer churn this quarter due to competitive and pricing pressures.” Technically, that’s about the same event. It just doesn’t tell a board member anything: is the problem pricing, onboarding, or account management? No way to know from that sentence. The fix is the instruction in the prompt above, plus a quick check that the specific nouns (Meridian, $3K, the champion who left) survived into the draft.
If drafting from scratch is where you get stuck most often, our guide on how to use AI to write a business report covers the same drafting mechanics in more depth, and most of it transfers directly to board memos.
A board memo’s job is to make the board smarter about your business in fifteen minutes. If the AI draft could describe any company in your industry, it isn’t done yet.
Every board has sat through a deck full of charts that looked fine and still left the room confused about what actually mattered. Sequoia calls this the calibration problem: it’s easy to add chart after chart, and much harder to pick the few correct ones that actually frame where the company stands (Sequoia Capital, Preparing a Board Deck). AI is genuinely strong at the first half of that problem, turning a spreadsheet export into a narrative. The second half, deciding which five or six numbers actually matter to this board this quarter, is a judgment call, and that part’s still on you.
“Here’s my metrics spreadsheet for this quarter: revenue, CAC, churn, burn, and headcount, each with the prior quarter for comparison. For each metric, write two sentences: what changed and why, based on the notes column. Flag anything that moved more than 15% in either direction. Do not round a specific cause into a vague one. If the notes say we lost a customer to a competitor’s pricing, say that, don’t write ‘competitive pressure.'”
That prompt works because it gives AI an explicit rule for handling the exact moment it wants to blur into generic language. It won’t decide which metrics belong in front of the board in the first place, though. That’s still your call to make.
Watch out for a second, quieter failure mode here too: AI will happily produce a confident-sounding explanation for a metric even when your notes column is thin or missing. That’s not the model lying to you, it’s doing exactly what it’s built to do: produce the statistically likely next sentence, whether or not your input actually supports it. If you didn’t actually write down why churn moved, don’t let the model invent a plausible-sounding reason to fill the gap. Ask it to flag anything it can’t explain from your notes as “cause unclear, needs input” instead of guessing. A board that catches you presenting a guessed cause as a known one will trust the next number you show them less, not more.
If most of your prep time goes into wrestling a spreadsheet into something readable in the first place, how to use AI to analyze a spreadsheet walks through the mechanics of that step on its own.
A number with no cause attached is just a data point sitting there. Give it a specific, named cause, the champion who left, the deal that slipped to a competitor’s price, and it turns into something your board can actually help you with. AI can draft that sentence. Whether the cause is the real one is still yours to own.
This is the step most people skip, and it’s the one that actually saves you in the room. Board Intelligence’s research recommends directors bring a standing set of questions into every meeting: what’s changed since last time, what assumptions the team is relying on, what’s being missed (Board Intelligence, How High-Performing Directors Prepare for Board Meetings). You can flip that same checklist around and use it on yourself before anyone else does.
“Act as a skeptical board member reviewing this memo and these metrics. Ask me the five hardest questions you’d actually ask in the meeting, the kind that target the weakest part of my story, not generic questions. For each one, tell me specifically which sentence or number in my memo prompted the question.”
A generic “what are your growth plans” question is close to useless, there’s no way to actually prep for that. What helps is a question that points at the exact sentence in your own memo that triggered it, because that tells you precisely where your story is thin. Run this two or three times with slightly different framing (a skeptical VC, a new board member, a director from a different industry) and you’ll start seeing the same three or four soft spots repeat. Those are the ones to actually fix before the meeting, not just rehearse an answer for.
This same rehearsal approach works for any high-stakes meeting, not just board sessions. Our guide on how to prep for any meeting in 15 minutes with AI covers a faster version for lower-stakes conversations, and ChatGPT prompts for managers has more variations on the mock-critic prompt if you want to adapt it beyond the boardroom.
If AI’s mock questions feel too easy to answer, they’re the wrong questions. Push it to be meaner. A board meeting is not the place to discover your first hard question live.
Back to that churn number. What happened to that founder has a name in AI research: sycophancy, a documented tendency for language models to drift toward agreeable, confident-sounding output rather than strictly accurate output, and the same pull shows up in summarization tasks, not just conversation (arXiv, Sycophancy in Large Language Models: Causes and Mitigations). That’s not a bug someone forgot to patch. It comes from how these models get tuned in the first place: human raters tend to score confident, tidy answers higher than messy, hedgy, specific ones, so the model learns to produce more of what gets rewarded. Ask AI to tighten a paragraph about a problem and it will often tighten the problem right out of the paragraph.
None of this is malicious. It’s just the path of least resistance for these models, since polished-and-vague is an easier combination to produce than polished-and-specific. Skip AI for this work and you lose all the time it saves you. Build one deliberate check into your process instead, the kind that catches the smoothing every time, and you keep the speed without the risk.
Here’s why this failure mode is so easy to miss: a smoothed sentence just looks like good writing, not like a mistake. In a board setting, that’s the dangerous part. The version that reads best sitting alone on a page is often the version that’s quietly hiding the most. A director skimming your memo won’t notice what’s missing. They’ll notice it live, in the meeting, when their question lands right on the gap the AI quietly closed.
Before anything goes to your board, read the AI-touched sections against your original raw notes, side by side. Anywhere the draft is smoother than your notes were, ask why. If the smoothing removed a number, a name, or a specific cause, put it back. This takes about ten minutes, and it’s the single highest-value ten minutes in the entire process.
The founder from the opening of this piece does one thing differently now. He keeps his rawest, ugliest notes open in a second window and checks every AI-smoothed sentence against them before anything goes near his board. Not glamorous. Maybe ten extra minutes of squinting at his own typos. But it’s the difference between a deck that survives questions and one that quietly falls apart under them.
Put the five steps on an actual calendar, not just a mental checklist. Sequoia’s advice to distribute board materials one to two days ahead, so the meeting itself is spent discussing rather than presenting, only works if your materials are genuinely ready by then (Sequoia Capital, Preparing a Board Deck). Separate research from NACD found 35% of directors say materials still arrive too late for proper review in the first place (National Association of Corporate Directors, Board Packs: The Elephant in the Boardroom), so the deadline discipline matters as much as anything AI helps you draft.
None of this needs to take longer than your current process. Most of it takes less time, because AI genuinely does remove the drudgery of turning notes into structure. What it doesn’t remove is the two human checkpoints this whole guide comes down to: deciding what the board actually needs to hear, and checking that the polished version still says it.
The same system works for investor updates between board meetings, just compressed. A monthly or quarterly investor email doesn’t need the full five-step timeline, but it still benefits from the same order of operations: collect your real numbers and notes first, let AI draft the structure, then run your own de-polish pass before you hit send. The version of this mistake that plays out over email is quieter than a bad board meeting, but it compounds. Investors remember which founders’ updates always sound great and never quite match what they hear elsewhere.
One takeaway if you remember nothing else from this guide: never let the last version of your board materials be one you haven’t personally checked against your own raw, unpolished notes. That check is the job AI can’t do for you. It’s the one that actually protects you in the room.
No. Let it draft structure and a first-pass narrative from your own notes and numbers, then rewrite the judgment calls yourself: what matters this quarter, what you’re actually asking the board for, and any number that needs a specific, named cause attached. Treat the AI draft as a skeleton, never as a finished deck.
Check your company’s data policy first, and use business-tier accounts (ChatGPT Team or Enterprise, Claude for Work, Gemini for Workspace) rather than free consumer accounts, since business tiers generally don’t train models on your inputs by default. Strip anything genuinely sensitive, like unannounced financing terms or personnel details tied to layoffs, before it goes into any general-purpose tool.
Honestly, the differences matter less than the process you run through whichever one you already pay for. Claude tends to handle long documents and nuanced instructions like “don’t soften this” a bit more reliably, ChatGPT has the widest file-handling and plugin ecosystem, and Gemini integrates cleanly if your team already lives in Google Sheets and Docs. Pick the one you already use and put your effort into the prompts, not the tool.
Give it an explicit rule against softening inside the prompt itself, something like “don’t round a specific cause into a vague one,” and then do a manual de-polish pass where you check every AI-smoothed sentence against your original raw notes. This is documented behavior in language models called sycophancy, a drift toward agreeable-sounding output, and it shows up especially in summarization tasks like turning notes into a memo.
Start collecting raw notes and metrics about a week out, and aim to have materials in your board’s hands one to two days before the meeting so they can actually read them instead of skimming during it. Build in the mock Q&A rehearsal and the de-polish pass with at least two to three days of runway, since both steps tend to surface things worth fixing before you send anything.
I researched this by going straight to primary sources rather than secondary roundups: Board Intelligence’s own research with Cambridge Judge Business School on board pack length and director reading time, the CEOWORLD Global Boardroom survey on executive AI adoption, Sequoia Capital’s published guidance to its own portfolio founders on board decks, NACD’s research on board pack timing, and the peer-reviewed arXiv survey on sycophancy in language models that underlies the over-polishing failure mode described above. Every stat and claim in this piece is sourced and linked below.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.