How to Draft an Investor Deck Narrative with an AI Model

Here is a prompt you can run today. It does three things deliberately: it forbids invention, it demands structure, and it asks the model to label its own confidence.

You are drafting a two-minute spoken opening for an investor deck.
Use only the notes and numbers between <raw></raw>. Do not add market
sizes, growth rates, competitor facts or customer names that are not
in the notes. If a required element is missing, write MISSING and say
what document would fill it.

<raw>
[paste your meeting notes, customer call notes, and a list of your own
metrics with their source document next to each one]
</raw>

Output exactly this structure:
1. Why now: 2 sentences, each ending with a [source: ...] tag pointing
   to the note or document it came from.
2. What we do: 1 sentence, no adjectives.
3. Who buys: 1 sentence naming the buyer role, not "businesses".
4. Proof: 3 bullet points, each a number from the notes with its source.
5. Risk we are pre-empting: 1 sentence.
6. The ask: 1 sentence.

Then list, separately, every claim you wanted to make but could not
source from the notes. Do not soften this list.

Why this shape works: the model is good at restructuring sentences you wrote and ruthless at spotting gaps when you ask it to. The "MISSING" instruction converts a hallucination risk into a to-do list. And the final section is the highest-value output: the claims the model reached for are usually the claims your deck is quietly making without evidence.

A worked example. Suppose your raw notes include: "Call with ops director at a logistics firm, said manual customs filing costs them about two days per shipment; signed a pilot with them covering three routes; our dashboard shows 41% of filing steps automated in the pilot; bank statement shows receipt of the pilot fee." A good run produces a proof section like "pilot live on three routes, 41% of filing steps automated (dashboard export), fee received (bank statement)," and a missing list containing "market size for customs software," "competitor pricing," and "time saved in hours," because none of those were in the notes. That list is your homework before the first investor read.

Keep the notes raw. Do not clean them into tidy sentences first, because the mess contains the specifics (the actual customer words, the actual number) that make the draft sound like a person rather than a pitch.

Run the draft through a removal pass with a simple decision rule: if you cannot point to the document that proves a sentence in under one minute, the sentence goes, or it gets rewritten as an opinion you own ("we believe", "our bet is"). Here is the checklist.

Financial projections deserve a separate rule because they sink more early decks than any other slide. A pitch-deck advisory that has reviewed hundreds of decks argues projections should usually be omitted at pre-seed and seed unless traction or an explicit investor request requires them, and, when included, every assumption must be defensible in 30 seconds without opening the data room (Waveup). That is a good test for your opening too: each number you say out loud should survive a follow-up question. If it cannot, the model wrote it from vibes, and you should strip it.

Also strip any claim you cannot repeat out loud without checking. Investors notice hesitation, and a single "I think that number is..." costs more credibility than a shorter deck. The J.P. Morgan pitch-deck guidance is blunt on the related risk: be careful of claiming a blue ocean, and lead with the metric your industry actually uses, such as lifetime value against acquisition cost for direct-to-consumer businesses or recurring revenue for B2B software (J.P. Morgan). The model will happily claim a blue ocean because it reads like confidence. Your audit pass should catch it.

If you want a second pair of eyes on the removal pass, a short consulting call can be a faster way to pressure-test the story than another draft. Quanta Consulting offers a free 30-minute discovery call to talk through goals and get a tailored recommendation (Quanta Consulting).

An AI model does not know your numbers. Every figure it produces either came from your paste or from its training data, and the second category is where decks go wrong. Treat the draft as a set of claims to be reconciled, not a document to be trusted.

The method is a one-row-per-number audit. For each figure, record four things: the number as written, the exact line in your notes it came from, the primary document (bank export, signed contract, CRM export, dashboard screenshot, filed financial statement), and the arithmetic check. Then ask a second question for every computed figure: does the calculation reproduce? If the draft says "revenue up 3x year over year," recompute from the two actual periods. If it says "average contract value of $18,000," multiply price by count and see whether the product matches total revenue.

Primary source discipline matters because audited or filed financials carry defined meanings. Public companies publish forward-looking statements with explicit safe-harbor language because projections are estimates subject to assumptions and risk (SEC example). Your startup does not need that boilerplate, but it needs the same intellectual honesty: a projection is not a fact, and a historical number should match the source system it came from. Where a real filing exists, for instance a public comparable's numbers, quote it from the filing rather than from memory (SEC filing).

A short verification script can automate the arithmetic half. This one parses a CSV of claims and flags any row whose stated value does not match a recomputed value:

import csv

with open("claims.csv") as f:
    for row in csv.DictReader(f):
        stated = float(row["stated_value"])
        recomputed = float(row["base"]) * float(row["multiplier"])
        flag = "CHECK" if abs(stated - recomputed) > 0.01 * max(1.0, stated) else "ok"
        print(f"{row['claim']:<45} {flag}")

Run it on every projection that involves a multiplication: contract value times count, price times volume, monthly times twelve. It will not catch a wrong input, but it catches the drift that happens when a model edits a chain of numbers and the arithmetic no longer holds. That drift is common and easy to miss when the prose reads well.

The drafting is the fast part. The loop that produces a deck you can defend is four passes:

1. Draft. Run the prompt with the raw notes and check the MISSING list.
2. Fill. Answer each missing item with a real document, or delete the sentence it belonged to.
3. Audit. Apply the removal checklist and the number reconciliation, one row per figure.
4. Read aloud. Time the opening against the two-minute target and cut anything you stumble on. Stumbles usually mark a claim you do not fully believe.

Two rounds of this loop will change the draft more than ten rounds of "make it more compelling." The language inflation you see in first drafts (adjectives, borrowed comparisons, confident market claims) is a model doing what it was trained to do. Your job is to be the editor with the spreadsheet open.

Do not use the deck to substitute for the business-model thinking either. The Business Model Canvas, developed by Alexander Osterwalder, exists precisely so the strategy is explicit before you dress it up: customer segments, value propositions, channels, customer relationships, revenue streams, key resources, key activities, key partnerships and cost structure (Wikipedia). Strategyzer's own description of the canvas frames the same nine blocks as the toolkit for designing a business model rather than a document (Strategyzer). Small business advisors use it as a pre-step: Duquesne University's SBDC calls the canvas a foundation for the business plan rather than a replacement, and notes it forces founders to answer tough questions before pitching (Duquesne SBDC).

That is the deeper insight AI drafting gives you. A model surfaces where your reasoning is thin faster than a human co-founder will, but only if you ask it to and only if you audit the result. The two-minute opening is a proof of understanding, not a summary of ambition.

If you are drafting several decks for different audiences, or running the same loop across a portfolio, it helps to have one API for text and image models so the prompt and the output format stay stable. Quanta Inference puts text, image, video, audio and embedding models behind one API key, OpenAI-compatible, prepaid and priced in US dollars (Quanta Inference). That matters less for one deck and more when the loop becomes part of how you prepare for a fundraise: the prompt, the audit script and the source list become a repeatable process, and the model doing the drafting is a variable you can swap without rewriting the workflow.

Whatever tool you use, the decision rule stays the same: the model writes the draft, you own every number in it, and a sentence without a one-minute source does not ship.