AI Grades Your Deck Before a Human Does. Here's How to Pass.

For three years the conversation about AI and business communication has been almost entirely one-sided. Founders, sales teams, and agencies asked how AI could help them make the deck: draft it faster, design it better, personalise it at scale. Almost nobody asked the more uncomfortable question. What happens when the person on the other side of the table is using AI too?

That question is no longer hypothetical. The investor reading your pitch, the procurement lead scoring your proposal, the buying committee evaluating your RFP: a growing share of them now run your document through a model before a human ever opens it. The document you send has two readers now. The first one is a machine, and it is far less forgiving than the human behind it.

The Other Side of the Table Went Agentic

The numbers on the receiving end moved faster than most senders noticed. 82% of VC firms now use AI for deal-sourcing research, and 85% of venture capitalists use AI to automate daily tasks, up from 76% a year ago. More than 70% of venture firms report using data analytics or AI in their sourcing and screening decisions, up from under 40% just four years earlier. AI screening and triage tools cut the partner hours spent on each inbound deal by 60 to 80 percent, which means the first pass on your pitch deck is increasingly not a partner at all. It is a model that decides whether a human ever sees slide two.

B2B procurement is on the same trajectory, only larger. Gartner projects that by 2028, 90% of B2B buying will be AI-agent intermediated, pushing more than $15 trillion of B2B spend through AI-agent exchanges. Buyers already complete roughly 70% of their journey before contacting a vendor, and when they do evaluate, they rank the written proposal (the document, not the demo) as the single most decisive factor, ahead of presentations, references, and proofs of concept.

Put those two trends together and the conclusion is hard to avoid. The deck or proposal you spent a week on is now being read first by software whose entire job is to filter it out. The bottleneck used to be earning a human's attention. Now it is surviving the machine that rations it.

What the Machine Actually Checks

It helps to know what the first reader is looking for, because it is remarkably consistent. SaaStr's AI deck analyzer has now graded more than 4,000 VC pitch decks against the profile of companies that actually raised, and the pattern in the scores is sobering. Most decks cluster between 55 and 70 out of 100: "fair," not fundable. The failures repeat across nearly every low-scoring deck:

Weak competitive differentiation. The model can tell within a few slides whether you have a real wedge or a feature list dressed up as a strategy. Vague "we're faster and cheaper" framing scores like the hundred other decks that said the same thing.

Market sizing without bottom-up logic. A top-down TAM number lifted from an analyst report is the single easiest thing for a model to flag, because it has seen thousands of them and none of them survive scrutiny. Bottom-up math (units, price, motion) reads as real because it can be checked.

Buried credentials. Team slides that hide the one fact that matters (the relevant exit, the domain depth, the unfair advantage) under a row of generic headshots. A human skims past it. A model scoring for "founder–market fit" simply records its absence.

None of these are design problems. They are substance problems, and the machine reader is specifically good at catching substance problems that a polished template would otherwise paper over. A deck can look immaculate and still score "fair," because the first reader is grading the argument, not the gradient.

The Sameness Penalty

Here is the twist that should worry anyone leaning hard on generation tools. In 2026, investors and their models pattern-recognise an AI-generated deck within about thirty seconds, by one analysis. The early drafts these tools produce look clean and confident, and that is exactly the problem: they carry the same structure, the same section order, and the same hedged language as every other deck the model has already seen that week. Fluency without specificity reads as noise.

This is the quiet penalty of generation-only workflows. The faster everyone drafts from the same prompts, the more uniform the output, and the more a screening model rewards the documents that break the pattern with something it cannot have generated itself: a real number, a named customer, a proprietary insight, a claim grounded in something only your company knows. The machine is not impressed that you used AI. It is filtering for the things AI cannot fabricate.

That should reframe what "good" means. Speed of generation is now table stakes. Worse, undifferentiated speed is a liability. The advantage has moved to whatever makes your document survive a sceptical, well-read, tireless first reader.

Two Readers, One Document

Even when a human does open the file, the engagement window is brutal and measurable. 81% of decks get started once they're opened, but the real cliff is early: investors who make it to slide four finish the deck 82% of the time. The first three slides decide almost everything. Personalisation moves the same needle: naming the investor and tailoring the framing lifts engagement by around 29%, and decks adapted to the reader's known preferences see lifts up to 47%.

So you are writing for two readers with different but aligned standards. The machine wants substance it can verify, differentiation it hasn't seen, and claims that don't contradict each other. The human, three slides in, wants the same things expressed clearly enough to keep reading. A document that passes the first reader almost always reads better to the second, because the discipline that satisfies a model (specificity, internal consistency, grounded claims) is the same discipline that holds a human's attention.

How to Pass the First Reader

The practical move is not to out-generate the machine. It is to put your document through the same kind of scrutiny before you send it that it will face after. Three things matter most.

Ground every claim in something only you know. The figures, the named references, the bottom-up market math, the specific outcome from a past engagement: these are the parts a screening model can't dismiss as boilerplate and can't have produced on its own. Generic confidence is the thing being filtered out; verifiable specificity is the thing being filtered for.

Check for contradiction before a model does. The fastest way to fail an AI first-pass is a number on slide nine that disagrees with the appendix, or a claim that nothing in your own materials supports. A machine reader cross-references every page as a matter of course. Your review process should too, ideally against your own knowledge base, so a flagged claim comes with the source it should have matched.

Treat review as a first-class step, not a final glance. This is the core of why critique, not generation, is where the leverage now sits. If the receiving end is running your deck through an expert model, sending it without running it through one of your own is choosing to be surprised. The right architecture is the one that mirrors the table you're sitting at: review agents that read every page the way a sceptical partner or procurement lead would, grounded in your brand, your data, and your past winning work, and surface the contradiction, the unsupported claim, or the undifferentiated slide while you can still fix it. The drafting buys back the hours. The review buys back the risk of failing a reader you never get to argue with.

The era when you only had to convince a person is ending. Increasingly, you have to convince the machine that decides whether a person ever sees your work at all. The teams that win the next few years won't be the ones who generate the most decks, or the prettiest. They'll be the ones whose documents were already graded, honestly and against everything the company knows, before they were ever sent.

Your deck has two readers now. Make sure the first one doesn't stop the second from ever looking.

The Lurio Team