Imagine a team that once spent a week preparing a presentation. With AI, it can produce the same presentation in an afternoon.

That sounds like progress. What happens next decides whether it is.

Maybe the team uses the extra time to interview customers, check its assumptions, or find out its recommendation is wrong. Maybe it produces five more presentations instead. The time saved is the same. The value created could be very different.

The warning I take from Cal Newport

That's the question I take from Cal Newport's discussion of Jerry Seinfeld and artificial intelligence. Newport argues that AI agents push knowledge work toward a “cognitive assembly line”: a stream of tasks we try to complete faster. Finishing more tasks feels productive, but that model of work isn't necessarily right, and it can be miserable. His answer is to be clear about what matters in our work, do the hard, valuable parts well, and use tools selectively. He takes Seinfeld's example as a reminder that efficiency isn't the goal; doing the work well means accepting the difficulty it requires. (Read the episode transcript)

I agree. For me, the practical question is which effort improves the outcome and which effort only consumes the day.

Not all difficulty is the same

A writer struggling to find the right argument may be developing an idea. A writer struggling to fix document formatting is probably losing time. Both are frustrating, but only one belongs at the center of the craft.

Treating every hard step as sacred protects plenty of bad processes. Treating every hard step as waste removes some of the experiences through which people learn to think. The challenge is knowing which kind of difficulty I'm facing.

The same tool can open or close the question

Take an analyst asked why customers are leaving. An AI tool can help organize interview notes, compare categories, or suggest competing explanations. That kind of use supports deeper investigation.

But a polished explanation that shows up before anyone checks the evidence can end the investigation too early. Its fluency can give the team permission to stop asking questions. Whether the tool expands the inquiry or shuts it down depends partly on what the person using it counts as a finished job.

My standard is simple. The analyst should be able to explain the evidence, defend the conclusion, and say what would change their mind. A beautiful report is useful only if it helps with those things.

What one person can't fix alone

Individual discipline can't solve all of this. Suppose a company rewards people for producing reports, answering messages, and keeping projects visibly active. Give that company faster tools and it has every reason to demand more reports, more messages, and more visible activity.

An employee might want to use the saved time for careful thought. Their manager may have already assigned it somewhere else.

So I think adopting AI should include an explicit decision about how to use the time it frees. A team could reserve part of it for verification, customer contact, uninterrupted thinking, or just a more sustainable workload. If nobody decides, the calendar will decide by default.

Assistance can make the work harder in a good way

I'd also resist the idea that working with AI has to weaken craft. A useful critic can make a writer face a weak argument. A counterexample can force an engineer to rethink a design. Used on purpose, assistance can make the intellectual work more demanding, because it exposes weaknesses that would otherwise go unnoticed.

That's a better goal than making everything feel effortless.

Three questions before adopting a tool

  • What result are we trying to improve?
  • Which decisions do we need to understand well enough to defend?
  • How will we use the time we save?

These questions make efficiency answer to a purpose.

For me, Seinfeld's example points to the value of knowing what makes the work good and accepting the hard effort it takes to get there. I can then judge a tool by whether it helps me do that work well, rather than how much faster it lets me finish.

A successful afternoon with AI might end with fewer documents, a rejected proposal, and one much better question.

We need a definition of productivity that counts that as a good day.