I accidentally built an AI that may put me out of business.

In doing research for my newest book, curiosity got the better of me, and I built an AI tool that does (part of) my job a bit worse, but infinitely faster than I can.

Haje Jan Kamps
4 min readOct 23, 2023


You may have read my popular Pitch Deck Teardown series, where brave founders send me their pitch decks, and I lovingly, caringly, and with a moderate amount of snark rip them apart so we can all learn what (not) to do when pitching to VCs. I like to think of myself as a storyteller. It’s a transferable, useful skill.

For that, and a number of other fun reasons, I have a huge library of pitch decks. People keep sending them to me, and I dump them all in a folder, ‘maybe I’ll need to reference it in the future,’ I think, and then very rarely do. They range from 3 to 130 (!) pages. In terms of funding rounds, they range from $150,000 to $700 million raised. Cumulatively, if all those decks were successful in raising funds, the companies would have raised around $175 billion. Suffice it to say, not all of those decks were successful in raising funds.

My book on the right. And, apparently, a little robot reading it. Because I guess robots need to learn stuff, too.

Using AI for Pitch Deck Research

In any case, I started a really ambitious project while doing research for my most recent book — a follow-up of sorts to Pitch Perfect — where I really got into AI. I ran the decks through a bunch of different AI models, to see if I could find patterns or observations I wouldn’t otherwise have made. I learned a lot, and I’m sure I’ll write more about that separately. But the point is, I was teaching this AI model a lot about pitching.

Experiment 1: Make a pitch deck. At first, I was curious if I could get the AI to make a great pitch deck. I failed — there is too much nuance to the storytelling, and the AI would need to understand VCs better than it currently does for that to work.

Experiment 2: Predict if a deck would be successful. Then I was wondered if I could get it to analyze a deck to predict success — both as a company and, well, for the current funding round. That was a partial success; a lot of the successful decks had things in common, at least.

None of these experiments were particularly helpful for the book I was writing. Then I had an epiphany: Wait, if I’d been able to identify what made some decks successful, isn’t that what I kind of do for a living?

Experiment 3: Give feedback on an existing deck, comparing it to my library of successful pitch decks. One thing led to another, and I discovered that yes, I was able to cobble together an algorithm that is able to give pretty damn good advice. Not on how to make a business successful, but on how it is similar or dissimilar from known decks that successfully raised funds. And from there, give advice on what is missing from a pitch.

The tricky part: How would you know whether the advice was any good? If only there was a way of just using the advice I was already giving, and training the AI on that, in order to give decent advice? Well, of course, there was. So I built that part of the engine, too, ensuring it’s linking the issues identified in the deck with specific chapters in the new book project. But… Wouldn’t it suck to have to tell people to buy a book and read a particular page? Surely, there’s better technology for that?

Putting it all together: The birth of Pitch.Guide

I managed to write a MidJourney prompt that creates stylistically distinct, really beautiful illustrations. They are used throughout Pitch.Guide to orange the place up a bit. And yes, MidJourney designed the Pitch.Guide logo as well.

And that’s how I came full circle: What started as research for the book ended up being a tool to help research the book, which then evolved into a separate tool again.

In a nutshell, if you submit the deck, my goofy-ass little low-code tool will take your pitch deck, analyze what is good or bad about it, and tell you, along with links to the relevant chapter in what no longer is a book, but turned into something far more detailed and community-like.

In the process, I created a workflow that is, I reckon, as good as me at identifying problems in a deck, and about 70% as helpful as me when it comes to helping my pitch coaching clients. The thing is: that help is expensive, and getting a good read on what the issues are with a deck is super valuable in itself.

Will it become as good at tearing apart decks as me? Maybe. If and when it does, perhaps it’s time for me to do yet another career change. Until that time, I hope it’s a somewhat helpful tool for founders who are struggling to figure out why their decks aren’t getting the response they are hoping for.

You can try it out here.



Haje Jan Kamps

Writer, startup pitch coach, enthusiastic dabbler in photography.