AI Made Being Wrong Cheap

The real advantage of AI for game founders is not building faster. It is finding the pivot before you burn the runway.

Photo by Aubrey Antles on Unsplash

Every week I talk with game founders, and the same moment keeps coming up. They have built something, the numbers are not there, and they already know they need to change direction. They do not ask me whether to pivot. They ask how long it will take to figure out what the pivot should be.

That question used to have a brutal answer: months.

I have lived it. In my early founder years, we would spend weeks on a single direction. We read reports, built spreadsheets, talked to a handful of users, and then argued about the little we had found. By the time we had something resembling an answer, we had burned real runway just to learn that we were pointed the wrong way. The pivot itself was usually correct. The cost of finding it was what hurt.

Here is the part I want you to sit with. AI has made being wrong cheaper. Not wrong after a year of production, after you have hired a team, built the game, soft-launched it, and discovered that nobody cares. Wrong before production, when you are still researching, comparing, testing, filtering, and changing your mind.

This is where I think most founders are reading the AI moment wrong. The headline everyone repeats is that you can build faster now. That is true. You can ship a prototype in a weekend. But building faster has a trap inside it. If you are building something nobody wants, building it faster just gets you to failure faster. The need to pivot does not go away because engineering got cheap.

What actually changes is the speed of the pivot. And the pivot was always bottlenecked on one thing: information. How fast can you research a market, slice the data, test an assumption, and see the next move clearly enough to commit? That bottleneck is what broke open.

Here is how I would use it if I were deciding what game to build today. The same method works for a browser game on Poki, a downloadable game on Steam, or a mobile title. The domain does not matter. The method does.

Fire bullets before cannonballs, and bullets just got cheap

There is a concept from Jim Collins and Morten Hansen in Great by Choice that I keep coming back to. I wrote about bullets, then cannonballs back in 2021, and AI has only made the idea sharper. They studied why some companies thrive in chaos while their rivals do not, and they found that the winners fired bullets before cannonballs. A bullet, in their words, is an empirical test that is low-cost, low-risk, and low-distraction. You fire small bullets to learn what actually hits. Only once you have a confirmed hit do you concentrate your resources and fire the cannonball.

Founders have always had one problem with this. Bullets were never as cheap as the theory wanted them to be. A real market test took weeks, so founders fired too few bullets, or they skipped straight to the cannonball on a hunch and called it conviction.

AI changes the economics of the bullet. Here is how I would actually fire them.

1. Fan out the research. Keep the conclusion, not the pile.

Say I am weighing a browser-based idle game on a platform like Poki or CrazyGames, where the audience runs into the tens of millions of monthly players. The old move was one research thread at a time. The new move is to run several at once.

I would ask, in parallel: how saturated is the idle genre on web right now, what mechanics are retaining players, what realistic revenue share and web ad eCPMs look like, how large the addressable audience actually is, and how the top three competitors monetize. Five bullets, fired together, back in minutes.

The skill is not asking more questions. The skill is refusing the wall of text. You want the conclusion, not the pile of raw pages. If what comes back is ten pages of notes, you have done it wrong. You want the one paragraph that tells you whether this direction is alive or dead.

2. Distrust every confident number.

Our industry runs on repeated benchmarks. “Day 1 retention has to be 50 percent.” “Hyper-casual is dead.” “Your CPI needs to be under fifty cents.” These get passed around until they feel like physics.

A lot of them are blog-lore. They were true for one genre, on one platform, in one year, and then everyone kept quoting them. The most useful thing I do with AI is not generate numbers. It is pressure-test them. Where did this benchmark come from? Is it still true in 2026? What was the sample? Make the model trace the claim back to its source and argue against it.

I learned this the hard way watching founders inherit a 2019 hyper-casual benchmark and build a 2026 game against it. The market had already moved. The number they trusted was stale. Do not bet a quarter of your runway on a statistic you have not cross-checked.

3. Make it attack your idea, not flatter it.

The lowest-value way to use AI is to ask it what to build. It will happily agree with you. A model that flatters you is worse than useless because it gives you the emotional feeling of progress without the information.

The high-value prompt is the opposite. Here is my concept: a merge game with a roguelike meta layer. Tell me why it fails. Give me the strongest argument that this audience will not care. Name the three games that already own this space and explain why my twist will not move retention.

Now you are getting somewhere. When the answer comes back with Travel Town and the other merge incumbents, and a clear case that a meta layer rarely fixes a saturated core loop, you may have saved yourself three months. Use AI as your sharpest critic, not your oracle.

4. Run the same question until you realize the filter is wrong.

This is the one that changed how I think. I was helping scope an underserved audience for a game. First pass, I asked for teen players. Too broad, useless. Second pass, I narrowed by genre. Still soft. Third pass, I realized the thing I had been getting wrong was not the search. It was the filter. The right cut was not age at all. It was platform behavior: who plays long sessions on web versus quick sessions on mobile?

Each pass corrected the question, not the answer. And here is what struck me: that is a pivot. A small one, happening in an afternoon, that used to take a research cycle of weeks. The pivot stopped being a dramatic event and became a thing I did four times before lunch. When the cost of changing your mind drops far enough, changing your mind stops being scary and starts being the work.

5. Build something you can actually look at.

You do not make good decisions staring at a chat window. I have AI build the artifact: a comparison table scoring six game concepts on market size, competition, build cost, and retention fit; a one-page dashboard; a rough mockup of the core screen.

The point is not that the artifact is perfect. The point is that it gives your judgment something to push against. The moment I can see six concepts ranked side by side, with the weak ones visibly weak, my feedback gets sharper and more specific. The format you evaluate in changes the quality of the call you make. A founder reviewing a real dashboard gives better notes than the same founder reading a paragraph.

6. Then take it to a human who has shipped.

None of this replaces an operator who has actually launched in your space. AI gives you breadth and a coherent first draft of a strategy. A real human catches the thing the model got confidently wrong.

I once ran a full analysis that looked airtight. Then I took it to someone who had shipped web games. In two minutes, he told me the platform would never feature that genre and my eCPM assumption was triple reality. The model did not know the platform politics, the featuring behavior, or the real monetization floor.

That is the division of labor. Use the model for breadth. Use the operator for the load-bearing facts.

The new failure mode is never firing the cannonball

Notice what none of this did. It did not tell me what to build. It made being wrong cheap, and it made the wrong answers show up fast. In a few days, I changed direction several times. Every change was an upgrade because each one ran on more information than the last.

But here is where the Collins discipline still bites. Cheap bullets tempt you to never fire the cannonball. You can research forever now. You can generate a hundred concepts, slice the market a hundred ways, and feel productive the entire time. Ask what Collins would ask: are you on an undisciplined pursuit of more?

At some point, you have to take the confirmed hit and put everything behind it. The information got cheaper. The conviction still has to be yours.

I think about Balatro, sold in the millions and built largely by one person. Or Vampire Survivors, a near-solo team and a price under five dollars. Those games are not examples of AI magically finding the answer. They are reminders that a small team does not need a giant organization behind it when the idea hits. AI does not guarantee that hit, but it gives small teams more shots before they run out of time.

That is the shape of the thing.

Final words

AI did not repeal the hard law of building. If you make something nobody wants, you still have to pivot, the same as founders always have. What it gave you is speed at the exact point that used to be slowest: the finding.

The pivot was never the problem. The time to find the pivot was the problem. That is the part that just got cheap.

If you are sitting on a game that is not working, you no longer have a runway excuse for taking a quarter to decide what comes next. You can fire ten bullets this week. Fire them.