3 min read

Building product with AI is playing Tetris at level 15

When an AI agent can prototype in hours what used to take weeks, the product problem stops being speed and becomes judgment: deciding where each new piece fits without breaking what's already in production.

As an AI PM you don't always choose which piece falls. Sometimes others choose it: a director blows up the roadmap with something that wasn't planned, and suddenly you have to scramble with something nobody saw coming.

That used to mean weeks of rearranging everything. With AI the pace changed: what took days is now prototyped in hours.

The pieces never stop falling

The backlog works just like Tetris. It doesn't matter how much you sort out today: tomorrow there are more pieces falling, and they don't ask if you're ready.

And they fall faster and faster. According to the Vercel Ship 2026 summary compiled by Digital Applied, half of Vercel's roughly six million daily deploys are now started by an AI agent, not a person.

What a gap looks like

Every feature you add without it fitting leaves a gap. Those gaps don't go away: they pile up as debt until one day nothing else fits and you lose the game.

You can spot them through three signs.

  1. It shipped and nobody updated the docs

    The product changed, and the only thing that explains it is still describing the old one.

  2. Nobody ended up owning it

    It works, nobody touches it, and the day it fails you have to find out whose it was.

  3. Edge cases were left for later

    They showed up in production, got noted down, and are still there. It's a list of gaps with a creation date.

The real skill

A good PM doesn't choose the pieces. They choose where and when they land.

With agents, that decision comes up more often and with less time. So it's worth having it ready beforehand, as a list and not as a hunch.

  1. Does this close a gap we already have or open a new one?

    A piece that doesn't fit is debt, even if it looks good in the demo.

    Passes if you can name the gap it closes, or say which part of the product it connects to.

  2. Can we test it small before going all in?

    If it's prototyped in hours, testing it with a few people is cheap. Not doing it is expensive.

    Passes if there's a version you can turn off without breaking anything.

  3. Who owns it once it goes to production?

    A piece without an owner is a gap with an expiration date.

    Passes if there's a name, not a team.

The trap: the bureaucracy that drags along

A mistake gets fixed fast. What doesn't get fixed fast is the bureaucracy that piles up on top of that mistake.

That one does bury the game: the process nobody dared remove, the approval that no longer approves anything, the document that exists because it existed.

Frequently asked questions

What does it mean for a piece not to fit?
That what was added to the product doesn't connect with what already exists: it doesn't close a problem you had, it has no owner, or it wasn't tested before committing. That piece leaves a gap, and gaps pile up as debt.
If AI makes everything faster, why doesn't the product improve on its own?
Because it speeds up building, not deciding. You can add more pieces in less time, and that makes it more urgent to decide which ones go in and where.
How many questions should I ask before approving a feature?
Three are enough: whether it closes or opens a gap, whether it can be tested small before going all in, and who owns it once it's in production.
Where does the Vercel figure come from?
From the Vercel Ship 2026 summary published by Digital Applied, which reports that half of Vercel's daily deploys are started by an agent. It's a secondary source; the link is below.

Sources

Let's talk

If your backlog is already filling up with gaps, tell me how it's set up and we'll look at it together.