The Friction AI Removes

The other day, my dishwasher broke.

I took a few pictures and asked AI to help me identify the problem, find the replacement part, tell me where to buy it, and share a video explaining how to fix it.

Let's replay the situation, but with a different ending.

I took the same pictures and asked the same question. But this time, AI suggested replacing the entire drainage system under my kitchen sink. It even recommended upgrading the pipes to gold and titanium because they would be more durable.

I sent the information to a plumber and asked them to implement the solution.

When the document looks complete

Now imagine something similar happening at work.

I'm working on a project and need to communicate an idea to someone with a different area of expertise.

I understand what I want to accomplish, but I'm less familiar with some of the technical, operational, or business implications.

I ask AI to help me prepare a document explaining my idea.

I review it. Most of it makes sense, and the rest looks reasonable.

So I share it with the person who will be responsible for implementing it.

The two situations seem quite similar to me. So why would I be more likely to question the plumber's recommendation than the document I just shared at work?

Perhaps because replacing my kitchen pipes with gold and titanium sounds obviously unreasonable. I know enough about my kitchen to recognize that something is wrong.

But when AI introduces a technical detail in an area I don't fully understand, I may not have the knowledge to recognize whether it is necessary, reasonable, or even relevant.

And because the document looks complete, I might assume that those details have already been considered.

The decisions hidden in the details

Suppose my original idea was to help users identify relevant information about a company.

AI expands it into a specification that includes API integrations, web crawling, a relevance score, and an interface for reviewing results.

All of these might be reasonable suggestions.

But did I actually request them?

Did I decide that relevance should be expressed as a percentage? Do I know what that percentage means? Did I consider whether web crawling is necessary?

Maybe not.

The problem isn't necessarily that AI produced incorrect information. It is that I may have shared decisions I never actually made, without realizing their consequences.

Now consider the person receiving the document.

They might reasonably assume that the implementation details represent requirements. They may spend time investigating them, estimating the work, or even implementing something I never intended to request.

Eventually, we may discover that the original idea was much simpler.

What friction disappeared?

Before AI, preparing a detailed document required more effort.

I had to articulate what I wanted, consider possible approaches, and sometimes confront questions I couldn't answer.

That friction wasn't always productive. Some of it was simply the effort of writing, organizing information, or explaining something clearly.

AI can remove much of that work, and that's valuable.

But some of the friction served another purpose.

It forced me to notice where my understanding ended and where someone else's expertise was needed.

When AI fills those gaps with plausible explanations and implementation details, I may no longer notice them.

The document becomes more complete without my understanding necessarily becoming more complete.

And the effort hasn't entirely disappeared. Some of it has moved to the person receiving the document, who now needs to distinguish what I actually want from what AI suggested.

Using AI to find the gaps

I don't think the solution is to stop using AI to prepare documents.

I find it extremely useful for exploring ideas, organizing thoughts, and identifying possibilities I hadn't considered.

But perhaps there is another way to use it.

Instead of only asking AI to complete a document, I could also ask it to identify the decisions embedded in that document.

Which requirements came from my original request?

Which implementation details are suggestions?

What assumptions have been made?

What would someone with a different area of expertise need to clarify before proceeding?

That would preserve something valuable about the friction AI removes.

I would still benefit from the speed and the exploration, but I would have a better chance of recognizing the limits of my own understanding.

Perhaps the most useful document isn't the one that answers every question.

It's the one that helps us identify which questions remain open, and who should be involved in answering them.