This month, we’re going to explore what might be the most counterintuitive practice in the age of AI: knowing when to slow down.

Hang on, slow down? Yes, bear with me here.

Let’s bring the debate that’s probably been going on in your team for some time to a head. Some of your colleagues say that with AI we should be building and shipping even faster, prototyping in hours, and that perhaps we don’t even need to write code at all any more: we should just let the models go full auto. They’re just that good now.

Others worry that this speed is creating quality problems, that we’re accumulating technical debt faster than we can pay it down, and that codebases are becoming patchworks of AI slop that nobody fully understands.

So, who is actually right?

To an extent, both are right, but I believe they’re talking past each other. The question isn’t whether to use AI for speed. It’s when.

We’ll start by looking at this debate through the lens of Daniel Kahneman’s System 1 and System 2 thinking, and why AI has made the slow phases of work more important, not less. Then we’ll examine the illusion of speed: the thought process around rework costs, and why going fast in the wrong phase means going slow overall.

We’ll explore when deliberate slowness pays off, including using AI itself for the slow work, and how fast prototyping is actually a form of slowing down. And finally, we’ll grapple with a question that’s getting harder to answer: why are you taking so long?

This article builds on themes from recent months. If you’d like to dig deeper, here are some related reads from the archive:

• One bottleneck at a time introduces the idea of subordination: telling the fast parts of a system to slow down so the constraint can catch up.
• Use it or lose it covers the thinking first protocol, a practical approach to slowing down before offloading work to AI, ensuring that critical skills aren’t lost.
• Invert, always invert explores pre-mortems and backward thinking, both examples of deliberate slowness in action.

Let’s get going.

Two speeds of thought

There’s a useful way to frame the debate that we opened with. In his oft-cited Thinking, Fast and Slow, Daniel Kahneman describes two modes of thinking: one that’s fast, automatic, and pattern-matching, and another that’s slow, deliberate, and analytical.

Transpose this thinking onto LLMs: in his conversation with Dwarkesh Patel, Andrej Karpathy describes them as ghosts or spirits, a kind of statistical distillation of human text, ethereal entities that are fully digital and mimicking humans. Words go in, patterns get matched, and words come out, which is, if you think about it, essentially, System 1 thinking.

AI is extraordinarily good at this kind of work: fast pattern-matching at scale. But the second kind of thinking, the work of deciding what to build, why it matters, and whether we’re solving the right problem, still requires human judgment.

And here’s the counterintuitive and highly interesting part: AI didn’t make the slow phases less important, it made them more important. When execution is cheap and fast, the leverage shifts to the decisions that precede it.

A wrong requirement, a misunderstood problem, a flawed design assumption: these propagate through everything AI helps you build, only now they propagate faster. The cost of getting System 2 wrong goes up precisely because System 1 has become so powerful.

If we want to go fast, we need to slow down first.

The illusion of speed

Back when I was doing my PhD, there was a common saying in academic circles: something like a few weeks in the lab will save you hours in the library. Software development has its own version: weeks of coding can save you hours of planning.

The reason it’s a joke, of course, is that both are backwards: the rush to start, the mounting realisation that something fundamental is wrong, and the painful rework that follows. I’ve certainly done many software projects where I wish I’d stopped and thought a little more before I rushed in. I can feel the cold flush coming back to me that you get when you stare at weeks of work that are completely wrong.

We have a clear intuition in software engineering that we should catch mistakes early, ideally in requirements or design, because the further a project moves on, the more expensive it is to fix them. Common sense can derive this without any research to back it up: a box diagram is easy to change, a misunderstood requirement less so, and a fundamentally flawed deployed architecture is a rewrite.

Therefore, here’s the problem: AI can help you create technical debt faster than ever! Oh no!

If the decisions that precede execution are flawed, AI will faithfully implement those flaws in a way that looks like fully featured code. Looks can often be deceiving, especially with powerful and confident models. It will generate thousands of lines of code based on a misunderstood requirement. It will happily build an elegant solution to the wrong problem.

The illusion of speed is that you’re making progress when you’re actually digging yourself into a deeper hole.

The answer isn’t to abandon speed, but to deploy it deliberately. We should only unleash AI’s pace when we’re confident it’s pointed in the right direction. Which raises the question: how do we know when that is?

When slowness pays off

The places where deliberate slowness pays off haven’t changed much, even as everything around them has accelerated. Requirements are still cheap to change when they’re just words on a page, and expensive when they’re deployed code serving real users. Design decisions are still easier to revise in a diagram than in a production system. AI didn’t alter this fundamental physics; it just increased the leverage of getting it right.

In a previous article, I called this the thinking first protocol: before offloading work to AI, spend time clarifying what you actually want. This isn’t unnecessary process; it’s the cheapest possible place to catch mistakes.

Here is the interesting paradox which shows the incredible usefulness of AI: the same tool that accelerates execution can also accelerate deliberation. Here are some practical ways to do this:

• Clarify requirements before coding. Spend 10 minutes writing down the problem you’re solving, your success criteria, and your constraints before asking AI to generate anything. What does “done” look like? What’s out of scope? Then get AI to interrogate everything that you’ve written before generating.
• Run a pre-mortem. Ask AI “What could go wrong with this approach?” before committing to a design. It will surface risks you hadn’t considered.
• Invert the problem. Ask AI “What would make this project fail?” to expose hidden assumptions. I’ve written more about this technique in Invert, always invert.
• Build a throwaway prototype. Use AI to create something in hours, show it to stakeholders, and validate your understanding before investing weeks. This is speed in service of slowness: you’re investing time upfront to learn.
• Build scrappy internal tools. Before you spend money on real products, use AI to build your own rough versions first. You’ll learn what you actually need and what you don’t. If you’re a paid subscriber, last month’s article goes deeper into some of the tools I’ve built myself.
• Surface edge cases early. Ask AI to generate edge cases and failure modes for your design before implementation begins. It’s far cheaper to handle them in a diagram than in production.

Of course, slowing down is easier said than done. Even if you’re convinced it’s the right approach, you’ll likely face resistance from those who see AI as a reason to speed up, not slow down.

The new cultural headwind

Given that AI is speeding things up so much, if you haven’t already been challenged on why something’s taking so long, you certainly will be soon.

“Can’t you just use AI?” is a new form of velocity pressure, and it’s particularly insidious because it conflates the appearance of productivity with actual throughput. Yes, AI can generate code in seconds. But generating code and solving the right problem are not the same thing.

So, what do you do?

• Be explicit about which phase you’re in. If you’re in the slow phase, say so: explain that you’re clarifying requirements, thinking through edge cases, and making sure you’re solving the right problem.
• Invite stakeholders to contribute. Their input is cheap to incorporate now and expensive later. Once you’re confident you’re pointed in the right direction, you can move fast.
• Show your working. Share artefacts from the slow phase: requirements docs, design sketches, pre-mortem outputs. This makes the invisible work visible and builds confidence that you’re progressing, not stalling.
• Timebox the slow phase. Give the slow phase a clear boundary: “We’ll spend two days clarifying requirements before we write any code.” This makes deliberate slowness feel intentional rather than open-ended.
• Share what you’re learning. Send brief updates as you discover things: edge cases you hadn’t considered, assumptions that turned out to be wrong. This turns the slow phase into a visible stream of value.
• Demonstrate quick wins. Build a throwaway prototype or mockup early to show stakeholders you can move fast when needed, buying you credibility for the slower, more deliberate work.

Interestingly, this maps nicely to the hill chart concept from Basecamp’s Shape Up methodology: the uphill climb is the slow phase of figuring things out, where uncertainty is high and you’re discovering what the work actually is; the downhill is the fast phase of execution, where the path is clear and you’re just getting it done.

This isn’t an excuse for delays; it’s a description of how good work actually gets done. The teams that ship fastest over the long term are often the ones that slow down at the right moments.

Your turn

This doesn’t have to wait for your next big project. You can apply this to every AI-assisted task you do. Before your next one, try this:

• Spend 10 minutes writing down what problem you’re actually solving. What does success look like? What’s out of scope?
• Before you start building, ask AI to run a pre-mortem on your approach. You might be surprised what it surfaces.
• If the task is significant, consider building a throwaway prototype first, one you’re willing to delete, just to validate that you’re headed in the right direction.

Wrapping up

Speed and slowness aren’t opposites; they’re tools for different phases. AI is effective for both: fast execution when the direction is clear, and accelerated deliberation when it isn’t. The skill is knowing which phase you’re in and applying the right tempo.

As always, until next time.