# Waterfall vs. Agile in a World of AI

> A good brief saves tokens, but it does not build the right thing. With AI, the loop got so fast that the back-and-forth is where the value is.
- **Author**: Nader Cserny
- **Published**: 2026-11-19
- **URL**: https://nader.io/posts/waterfall-vs-agile-in-a-world-of-ai

---

<img src="/posts/waterfall-vs-agile-in-a-world-of-ai.webp" alt="Drawing on black: grey bars step down like a waterfall and end at a grey cross, captioned 'Long upfront specs faithfully embed wrong assumptions.' In the middle, a green line loops in tightening circles to a bright green dot. On the right, a green label reads 'The back-and-forth is the point.' above the line 'Saving tokens is nice. Building the right thing is the job.'" class="post-hero" width="1200" height="630" />

A chief product officer wrote on LinkedIn recently that briefing AI well matters more than ever, because a good brief saves tokens. I agree with the part about tokens. I disagree with where it leads.

Yes, a tight brief is cheaper: fewer turns, less context, fewer retries. If the token bill is what you're optimising, write everything down first and let the model run.

But that's Waterfall. Working with AI makes the old Waterfall vs. Agile debate matter more, not less.

## The big upfront brief

I tried it. I had AI write a full product brief for a data visualization tool, covering the features, user flows and components. Then I handed the document to another AI and let it build.

It sort of worked. It implemented what was in the document, but the result was still far from what I had in mind. I redid a lot of components and reworked most of the user flows. Every gap and wrong assumption in the brief got built faithfully, and nobody asked a question along the way.

If I had started with a small MVP and iterated from there, I would have got there faster and spent less overall, tokens included.

## Agile got faster, not obsolete

The Agile loop still holds: ship an MVP quickly, look at it, gather feedback, iterate. What AI changes is the speed. A loop that took a sprint now takes minutes. Waterfall's best argument used to be that iterating was expensive. That argument is mostly gone. What's left is the token bill, and it's small next to an afternoon of rework.

So a massive spec, probably written by AI and handed to another AI that builds all of it with no milestones and no questions, is the worst of both worlds. You pay for the detail and still get the wrong thing.

## The back-and-forth is the point

For me, back-and-forth is simple. I prompt, the AI comes back with something, and I can see straight away whether I like it. That reaction is the useful part. Seeing an actual version gives me ideas I would never have had staring at a blank document.

Sometimes the AI surprises me, too. It comes up with something more creative than what I had in mind. A long spec leaves no room for that. At best, you get exactly what you asked for.

## Where Waterfall still wins

There are exceptions. We once trained a model to recognise people and companies in text. A job like that needs a lot of data and a lot of definition up front, because a single training run takes hours. When each loop is that slow and that expensive, you want to get as much right as you can before you press go.

That gives a rule of thumb: the more one loop costs, the more you should plan before it. For most product work with AI, a loop is cheap.

## Brief well, then talk

None of this means briefing doesn't matter. Give the AI context, constraints and a picture of what good looks like. Just treat the brief as the start of a conversation, not a replacement for one. Set milestones, look at what comes back, and steer.

Saving tokens is nice. Building the right thing is the job.

---
- [All articles](https://nader.io/posts)